Internet-Draft Agent Coordination in Optical Networks November 2026
Yang, et al. Expires 20 May 2027 [Page]
Workgroup:
CCAMP Working Group
Internet-Draft:
draft-yang-ccamp-optical-agent-coordination-00
Published:
Intended Status:
Informational
Expires:
Authors:
H. Yang
Beijing University of Posts and Telecommunications
M. Han
Beijing University of Posts and Telecommunications
Q. Yao
Beijing University of Posts and Telecommunications
T. Yu
Beijing University of Posts and Telecommunications

An Architecture for Agent Coordination in Computing-Aware Optical Networks

Abstract

Computing-aware optical networks coordinate optical connectivity with the capabilities, availability, and performance of computing services. This document describes an informational architecture for coordination among joint control, transport control, network element, and compute resource agents. These roles relate local observations to domain abstractions, assess connection-and-site candidates against a common service objective, and evaluate the combined service outcome. Computing resources may be located at edge, regional, or centralized sites. Coordination supports access to existing service instances and, where authorized, new workload placement or optical resource allocation. Resource managers retain admission and execution authority. The architecture is independent of optical switching technology and agent algorithm. An Optical Transport Network (OTN) and edge computing provide a detailed realization example; reinforcement learning is an optional policy use case. This document defines no new protocol, YANG module, or optical signal format.

Status of This Memo

This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.

Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.

Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."

This Internet-Draft will expire on 20 May 2027.

▲

Table of Contents

1. Introduction

Optical connectivity and computing resources jointly determine whether a service can meet its objectives. A low-delay optical connection may terminate at a busy compute site, while an available site may be reachable only over a connection that cannot meet the capacity, protection, or activation-time requirements. Selecting a site changes the required network endpoint; available connectivity constrains which sites can be used. Independent network and compute decisions can therefore produce a combination that does not meet the end-to-end service objective.

In this document, computing awareness means that service coordination uses service-specific compute capabilities and current availability together with optical connectivity and feasibility. The relevant knowledge is exchanged through the logical roles that need it; individual optical elements need not maintain a complete compute-resource view. Awareness informs a decision, while authorized resource managers determine whether that decision can be admitted and executed.

This document describes cooperation among control agents, network element agents, and compute resource agents. Local observations and capabilities inform domain abstractions, while joint control evaluates the coupled service objective across transport and compute resources. The architecture identifies which view informs each decision, which authority can change resources, and how resulting observations inform service assurance. Its roles can be realized by different agent implementations within existing control and resource-management systems.

The architecture uses Abstraction and Control of TE Networks (ACTN) control and abstraction functions [RFC8453], their information model [RFC8454], and TE topology representation [RFC8795]. OTN terms follow ITU-T G.709/Y.1331 [G709]. Computing-Aware Traffic Steering (CATS) provides network and compute inputs for service-site selection [RFC10053]; its OTN applicability is discussed in [I-D.zhao-cats-otn-applicability]. The agent roles in this document augment existing control and resource-management functions.

Related individual Internet-Drafts discuss network management agents, agents within ACTN optical controllers, and optical-compute orchestration [I-D.zhao-nmop-network-management-agent] [I-D.zhao-ccamp-actn-optical-network-agent] [I-D.tan-ccamp-onco-control-framework]. This document proposes an architectural description of the relationship between local intelligence, domain abstractions, and joint service decisions. The scope and its relationship to those drafts require community review.

A service's semantic profile is a bounded, operator-defined description of its objectives. It can supply defaults within the limits of an authorized request but cannot grant priority or relax hard constraints. Agent cooperation does not imply payload inspection, unrestricted natural-language commands to equipment, or dependence on a particular AI model.

2. Terminology

Computing-aware optical network
An optical network whose service coordination considers computing-service capabilities, availability, and performance together with transport connectivity and feasibility.
Service intent
An authorized request containing service objectives, endpoints, transport and compute needs, and policy context.
Semantic profile
An operator-defined behavior and objective profile, such as latency-sensitive or bandwidth-intensive, that can supply defaults within operator-defined limits.
Agent
A logical function that observes relevant state, reasons about an assigned objective, selects permitted actions, and evaluates outcomes within an explicit authority boundary. It may maintain internal context; no model family or task framework is prescribed.
Control agent
An agent hosted by or associated with a controller or service coordinator that uses a domain or service view to make decisions and coordinates decisions within that function's authority.
Joint control agent (JCA)
A control-agent role that coordinates optical connectivity and the selection or use of computing resources against one service intent. It implements the joint coordinator described in this document.
Transport control agent (TCA)
A control-agent role that relates element capabilities and observations to a transport-domain view and uses the domain controller's resource-control functions.
Network element agent (NEA)
An agent embedded in or adjacent to an optical network element that uses local capabilities, constraints, and observations to make decisions, contributes to the domain view, and acts within owner-authorized local control.
Compute resource agent (CRA)
An agent within or associated with a compute resource manager that assesses eligible sites, service readiness, usable capacity, and permitted resource operations. It can support access to an existing service instance or authorized workload placement.
Optical transport domain controller
The authoritative function for domain topology, connection feasibility, reservation, provisioning, and reconciliation; an ACTN PNC can perform this role.
Joint coordinator
The logical service coordination function, implemented here by the JCA, that evaluates path-and-site pairs and correlates their lifecycles.
Task context
The correlation, objective, scope, policy, deadline, and evidence information needed to interpret and safely continue an assigned task.
Delegated action
An action permitted by the resource owner for specified resources, operations, and validity. An agent's proposal alone is not a delegation or resource reservation.
MDSC
Multi-Domain Service Coordinator in the ACTN architecture.
PNC
Provisioning Network Controller in the ACTN architecture.
AI
Artificial intelligence.
DQN
Deep Q-Network, a reinforcement learning method illustrated as a local policy tool in this document.

3. Scope and Design Objectives

The architecture concerns functional cooperation across service coordination, transport-domain control, network element intelligence, and compute-resource management. Its scope is the abstraction of capabilities and constraints, the coupling of service objectives, the allocation of decision authority, and feedback across control levels. Logical roles may be combined or distributed in a deployment.

The optical network provides connectivity between the service ingress and eligible computing sites. Computing resources provide the service processing capacity, whether at an edge, regional, or centralized site. Joint coordination evaluates their combined contribution to the service. Selecting a site can select an existing service instance or request a new workload placement; the compute resource manager determines the applicable admission and placement operations. This architecture concerns the relationship between these resource domains rather than a particular compute location or optical switching technology.

The architectural objective is that a joint service decision is consistent with authorized objectives, currently available domain capabilities, and confirmed resource states. The relationships remain applicable when resource selection, admission, and provisioning are implemented by different workflows. Local agents can use local knowledge to refine a domain abstraction or trigger service review within their authority boundaries.

This document is Informational. It describes proposed functional roles and architectural relationships. It specifies no interoperable encoding, agent-to-agent protocol, new ACTN interface, YANG module, mandatory task schema, or provisioning state machine. Detailed workflows in Appendix B illustrate one realization; they do not prescribe the architecture. A later interface specification would require an explicit interoperability gap analysis.

Agents can use rules, optimization, reinforcement learning, or model-assisted reasoning to realize these functions. No learning method, large language model, training reward, autonomous-network level, or internal task decomposition is required. The architectural description concerns the information, decisions, and authority visible across functional boundaries.

4. Architectural Principles and Abstraction Boundaries

The architecture is described through objectives, views, decisions, and authority. A deployment can realize these relationships using centralized orchestration, distributed cooperation, pre-established resources, or event-driven control. An abstract relationship does not establish a particular message sequence or require a separately deployed agent process.

4.1. Coupled Service Objectives

A service objective spans transport reachability, compute-site eligibility, and the resulting service behavior. Transport and compute choices are coupled because the chosen site determines the required connectivity and the available connectivity limits usable sites. Authorized hard constraints bound the joint decision; local optimization can express preferences within that feasible region.

A candidate combines an ingress-to-site connection with a compute site and its service or workload capability. It is usable only when both domains confirm their contribution against compatible service constraints. The transport view includes the connection endpoints, feasible capacity, delay, protection, and activation conditions. The compute view includes service readiness, usable capacity, processing-delay estimates, locality, and placement conditions. Joint control checks the combined service budget and the consistency of these conditions before selecting or retaining a candidate.

Coordination is bidirectional: a compute-capacity change can require another connection or site, and an optical resource change can make a previously usable compute site unreachable under the service constraints. The JCA can request revised domain assessments or alternative candidates when either view changes. The architecture allows iterative or concurrent assessment; it does not require the network to be selected before the computing resources, or the reverse.

A realization may select a new connection and placement, retain an existing service, adjust resources within the current domain, or reject an objective that cannot be met. Joint coordination therefore applies to service feasibility and assurance as well as initial resource selection.

4.2. Layered Resource Abstraction

The service view describes objectives and acceptable outcomes. Domain views describe capabilities exposed under operator policy, constraints, feasibility, and observed resource states. The local view includes equipment or workload details used to form those domain views. Each view exposes sufficient information for the decisions at that level while retaining technology-specific state and confidential detail with the responsible owner.

Abstraction preserves decision-relevant constraints. An abstract capacity or delay value is accompanied by its applicability, validity, and feasibility conditions where needed. A local change can invalidate a domain feasibility assessment, but it does not grant a local agent authority over another domain. The joint control function uses compatible domain views to make decisions without requiring all internal resource details.

4.3. Knowledge, Decision, and Authority

An observation or inference contributes knowledge; a proposal expresses a possible decision; a resource state confirmed by the responsible resource manager establishes what was actually allocated or changed. These categories remain distinct regardless of the algorithm or coordination mechanism. Joint assurance relates confirmed resource states to observed service outcomes.

Authority follows the resource-control boundary. Local and domain autonomy operate within delegated policy, while decisions that affect the combined service are reconciled at the joint coordination level. Control levels can operate at different timescales; local protection remains effective independently of higher-level reasoning.

5. Operational Problems and Design Properties

Joint selection accounts for the interaction between transport and compute constraints. A service request can arrive before a suitable optical connection exists, while the corresponding workload placement may be owned by a different controller. Network metrics, compute occupancy, and the state of an optical cross-connect can change between candidate selection and actual activation. The following properties describe the problems that an implementation of this architecture needs to address. They are architectural considerations rather than new protocol requirements.

Table 1: Operational problems and associated design properties
Problem Design property Observable result
Untrusted or ambiguous intent Authenticate the requester, check authorization, and validate sites, profile defaults, units, and hard constraints before selection Rejected request or recorded normalized constraints
Abstract capacity alone does not establish feasibility Ask the optical transport owner to verify technology-specific feasibility, connectivity, capacity, switching, and protection Feasibility result with validity and reason
Separate resource owners Correlate time-limited reservations and confirm each activation Both resources active or partial resources released
Concurrent allocation Treat feasibility as a snapshot and the hold as the resource decision Failed hold triggers cleanup and a fresh decision
Unequal observation times Carry timestamps and validity for transport and compute state Stale data cannot justify hard-constraint admission
Disruptive re-optimization Account for setup time, affinity, protection, overlap capacity, and change limits Change only when the service policy permits it

These properties also set a boundary for optional optimization algorithms. A ranking function can propose a candidate but cannot authorize a requester, make unavailable transport resources available, or make a partial activation successful. Keeping admission and lifecycle checks outside the ranking function makes it possible to replace a heuristic with a learned policy without changing the meaning of a service request.

Agent coordination also addresses disagreement between a local observation and a domain view, and incompatible proposals affecting the same service. The TCA reconciles local reports with authoritative transport state; the JCA checks transport and compute results against a common objective, relevant resource scope, and compatible validity. An NEA can withdraw an earlier assessment without gaining authority to select a compute site or allocate another domain's resources. Late or superseded proposals do not authorize resource changes.

7. Agent Cooperation Architecture

 +----------------------------------------------------------------+
 | Layer 1: Service Intent                                        |
 | Authorized objectives, constraints, and acceptable outcomes    |
 +-------------------------------+--------------------------------+
                                 ^
                       objectives / policy
                       service outcomes
                                 v
 +----------------------------------------------------------------+
 | Layer 2: Agent Cooperation and Control                         |
 |  +----------------------------------------------------------+  |
 |  | Joint control agent (JCA)                                |  |
 |  | Service coordinator / ACTN MDSC                          |  |
 |  | Joint path/site decisions and service assurance          |  |
 |  +------------+-------------------------------+-------------+  |
 |               ^                               ^                |
 |       objectives / views              objectives / views       |
 |               v                               v                |
 |  +------------+-------------+    +------------+-------------+  |
 |  | Transport control agent  |    | Compute resource agent   |  |
 |  | (TCA): domain view       |    | (CRA): compute view      |  |
 |  | Optical controller / PNC |    | Compute resource manager |  |
 |  | Optical resource control |    | Compute resource control |  |
 |  +------------+-------------+    +------------+-------------+  |
 +---------------+-------------------------------+----------------+
                 ^                               ^
       policy / permissions              policy / permissions
       local views / outcomes          local views / outcomes
                 v                               v
 +---------------+-------------------------------+----------------+
 | Layer 3: Resources and Execution                               |
 |  +------------+-------------+    +------------+-------------+  |
 |  | Optical network elements |    | Computing sites          |  |
 |  | NEAs at elements X and Y |    | Edge, regional, central  |  |
 |  | Local views and control  |    | Execution observations   |  |
 |  +--------------------------+    +--------------------------+  |
 +----------------------------------------------------------------+

 Objectives and policy flow down; views and outcomes flow up.
 Resource owners retain admission and execution authority.
 Logical layers; functions may be combined or distributed.
Figure 1: Three-layer architecture for joint optical and computing resource coordination

Figure 1 groups the architecture into three logical layers. The Service Intent Layer expresses authorized objectives, constraints, and acceptable outcomes. The Agent Cooperation and Control Layer contains the JCA in a service coordinator or ACTN MDSC, the TCA associated with an optical domain controller, such as an ACTN PNC, and the CRA associated with the compute resource manager. The JCA coordinates joint service decisions across the transport and compute views; the TCA and CRA retain their domain responsibilities and resource-owner boundaries. The Resources and Execution Layer contains optical network elements with local NEA functions and computing sites. It supplies local capabilities, constraints, and observed resource states and performs owner-authorized actions.

The three layers are functional groupings rather than additional ACTN controller types. The JCA-to-TCA and JCA-to-CRA relationships are internal to the cooperation layer; grouping them does not remove the service-to-domain abstraction or their distinct decision authority. NEAs remain associated with local resources in the execution layer. Deployments can combine or distribute functions while preserving these relationships.

NEAs contribute local capability, configuration, operational knowledge, and bounded control. Their analysis can refine a domain view, identify an invalid feasibility assumption, or support recovery. The TCA reconciles this knowledge with the PNC's authoritative state. The element-to-domain relationship supports continuous observation and event-driven cooperation as well as requested assessments.

Cooperation couples three architectural relationships: local intelligence to domain abstraction; transport and compute views to a joint service decision; and confirmed resource states to service assurance. The downward relations convey authorized objectives or control policy; the upward relations convey abstract capabilities, constraints, and observed outcomes. They do not define an agent task protocol.

Coordination depends on the information contributed by each role and how it changes a service decision. The TCA reconciles NEA observations with transport-controller state, the CRA assesses usable computing services, and the JCA relates their results to the same objective. A local constraint can invalidate a domain candidate; a changed domain view can require a revised joint decision. The realization identifies these dependencies and the feedback used to review a decision, including when no resource change is needed.

Existing TE topology provides an abstract transport view [RFC8795], with OTN-specific representation under development [I-D.ietf-ccamp-otn-topo-yang]. Agent reasoning uses these views and owner capabilities. It augments the available knowledge and coordination while retaining the established topology and resource models.

7.1. Agent Roles and Resource Authority

Table 3: Agent responsibilities and decision boundaries
Role Architectural responsibility Execution boundary
JCA Service-objective reasoning, joint feasibility, combined assurance Coordinates owners; cannot override domain admission
TCA Element knowledge abstraction, domain feasibility, transport assurance Invokes PNC-authorized computation and operations
NEA Local capability and constraint reasoning, bounded local control Only delegated element operations; no global allocation
CRA Compute capability abstraction, placement feasibility, outcome analysis Invokes compute-manager-authorized resource operations

The PNC and compute manager retain their resource records, admission logic, and execution authority. Agent knowledge supports decisions but does not replace those authoritative records. Owner policy governs the relationship between a proposal and an actual resource change, including when a previously acceptable proposal has become invalid.

An NEA may be embedded in a device or run adjacent to it within the transport owner's trust domain. Its architectural role is local reasoning and bounded control informed by equipment context. A deployment documents the capability and authority associated with that role; deployment location and model choice remain implementation decisions.

7.2. Knowledge and Service-Assurance Relationships

Local knowledge informs domain representations, domain capabilities inform joint decisions, and observed changes in resource state inform assurance. Authorized policy constrains reasoning and control at each level. These relationships may be realized by continuous telemetry, event notifications, requested analysis, or internal information sharing.

Consistency is judged against the service objective and the validity of each contributing view. Local observations can update an abstraction without forcing an immediate service change. A domain exception becomes a joint concern when it affects combined feasibility or the observed service outcome.

Assurance closes the relation between intended and observed behavior. The architecture supports retaining, adapting, or withdrawing a service according to authorized policy. Tool invocation, internal planning loops, and detailed resource reconciliation are realization mechanisms illustrated in Appendix B.

7.3. Delegation, Conflict Resolution, and Control Timescales

The owner delegates operations for identified resources and a limited time interval. Delegation specifies which changes can proceed automatically and which require another owner or operator decision. Task delegation cannot enlarge this authorization. A local NEA can perform a pre-authorized action or report an exception, but a new route, compute-site move, or resource commitment requires the appropriate controller and joint service checks.

The TCA resolves conflicting NEA reports using source identity, observation time, resource revision, and authoritative device or controller queries. The JCA resolves conflicting transport and compute plans against the service's hard constraints. It does not accept the first favorable response when another required participant reports a constraint violation. Repeated contradictions cause an explicit failure response or escalation.

Equipment protection and other time-critical mechanisms continue under their existing control. Agent-assisted diagnosis and cross-domain replanning operate at a timescale suitable for their observations and actions. Local restoration is reported upward so that the JCA can distinguish a protected transport change from a failure that requires compute relocation. The architecture assigns one authoritative writer per affected resource during an execution attempt.

8. Logical Relationships and Information Abstraction

The architecture identifies logical relationships by the information and authority they convey. It does not prescribe peer discovery, transport bindings, task objects, or operation order. A relationship can be internal to one system or cross an authenticated management boundary while retaining the responsible resource owner.

8.1. Service-to-Domain Relationship

Joint control expresses authorized service objectives and interprets domain capability, feasibility, and outcome views. Transport and compute domains expose only the information required to evaluate their contribution to the same service. Comparable scope, applicability, and validity allow their views to support a coupled decision.

8.2. Domain-to-Element Relationship

Domain control relates local capabilities, constraints, and observed resource states to domain-wide connectivity and resource conditions. Network element agents contribute knowledge and permitted local control. The domain reconciles local information with authoritative state and policy before exposing a feasibility claim or changing resources.

8.3. Architectural Information Classes

Cooperation uses five conceptual information classes: service objectives and policy; capabilities and constraints; observations and their validity; decision proposals and affected scope; and owner-confirmed resource states and service outcomes. These classes describe semantics visible across functional boundaries, not a mandatory message schema.

A realization maps these classes to existing topology, service, telemetry, and management records where sufficient. It documents any remaining semantic gap before proposing new interface fields. Service correlation supports consistency across views; task identifiers, attempt numbers, and reservation tokens are optional realization-specific records illustrated in Appendix B.

9. Information Used for Joint Selection

9.1. Service Intent

Useful inputs include ingress and egress attachment points, a service identifier, requested bandwidth, latency and availability objectives, expected service duration, compute requirements, locality restrictions, and an optional semantic profile. A profile such as "interactive" is an operator-defined hint that can propose defaults; explicit authorized constraints take precedence. A requester cannot gain elevated priority or consume extra resources merely by claiming a sensitive class. The following table lists conceptual information, not a new encoding or registry.

Table 4: Illustrative service-intent information
Element Meaning and example unit Authority
Service and request identifiers Stable identifiers for correlation and retry Requester and operator
Ingress and eligible sites Attachment point and permitted compute locations Operator policy
Semantic profile Named behavior hint and profile version Operator after request validation
Transport constraints Minimum bandwidth in bit/s; maximum transport or end-to-end delay in ms; protection need Authorized service contract
Compute constraints Workload capacity, execution delay estimate, locality restrictions Compute resource manager
Validity Start time, duration, and freshness deadline Requester and operator policy

The translator records which defaults were derived from a profile and which constraints were explicitly requested. If constraints conflict, the request is rejected or returned for correction; silently weakening a hard constraint would change the requested service. Service identity should be mapped to existing service identifiers in the applicable environment rather than introducing another identifier in this architecture.

The coordinator first resolves the requester's identity and entitlement, then applies operator policy to any semantic profile. The normalized result distinguishes hard constraints from soft objectives. Hard constraints include permitted ingress and sites, a minimum admitted client capacity, required adaptation and protection, maximum combined delay, placement restrictions, and any latest acceptable activation time. Soft objectives can include preferred site, lower estimated delay, or lower resource cost. An absent bound is not interpreted as an unlimited authorization: operator policy supplies a bounded default or rejects the request. Units, the measured portion of a delay bound, observation time, and the source of each value need to be unambiguous within the deployment.

For a delay objective, the operator defines the measurement endpoints and whether the value is one-way, round-trip, or a service transaction time. A transport-delay estimate and an execution-delay estimate can be added only when they cover non-overlapping parts of that same objective and use compatible statistical meaning. Queueing, access, and application processing outside those estimates require an explicit budget or margin. An estimate is not a guarantee; when a hard bound cannot be checked with the available information, the coordinator does not declare the pair feasible. A bandwidth request similarly needs a defined client rate and adaptation context before the optical domain controller can determine the applicable reservable resources; client adaptation is part of that mapping where required by the transport technology.

When CATS is used, the service identifier can be the CATS Service ID (CS-ID), and a candidate contact point can be identified by its CATS Service Contact Instance ID (CSCI-ID) [RFC10053]. The request identifier used for setup retries is a separate operational correlation value, not a replacement for CS-ID. The coordinator maps a chosen contact point to an optical transport egress attachment under operator policy; a CSCI-ID alone does not prove that a client signal can be adapted or connected at that egress.

9.2. Transport and Compute State

The transport view can expose candidate connectivity, residual capacity, latency estimates, protection characteristics, and fault state at the abstraction level authorized for the coordinator. The compute view can expose eligible sites, available capacity, placement restrictions, and estimated execution delay. Measurements need timestamps and validity intervals so that decisions do not treat stale telemetry as current capacity.

Compute capability and load are interpreted for the requested service. A utilization percentage or server count does not, by itself, establish equivalent service capacity across sites. The CRA identifies the applicable service or workload profile, capacity and delay meanings, supported execution requirements, and observation validity. Candidate comparisons use compatible definitions or a documented mapping; a ranking score is kept distinct from an admissible capacity or delay bound. This document defines no universal compute unit or normalization function.

The optical domain interprets these abstractions using its own switching technology. In a wavelength-switched network, routing and wavelength-assignment feasibility depends on wavelength availability, node connectivity, and conversion capabilities [RFC6163]. In an OTN realization, feasibility depends on the applicable client adaptation, multiplexing, and switching constraints. These are alternative technology-specific checks; they are not universal resource units or checks that every optical domain must perform. The JCA uses the resulting feasible connectivity view without performing the domain-internal assignment itself.

For OTN connection admission, the transport domain also considers the supported client-signal adaptation, the connection endpoints, switching capability, reservable transport units or tributary resources, protection relationships, and provisioning time. An abstract bandwidth value alone may overstate usable capacity when resource granularity, continuity, or cross-connect constraints prevent a particular request. The coordinator therefore treats a domain-controller feasibility response as authoritative and does not infer optical feasibility from packet queue occupancy.

For this OTN realization, a useful feasibility response identifies the candidate's ingress and egress, whether the requested client signal and adaptation are supported, whether the requested capacity and protection can be reserved, the estimated transport delay and setup time, the validity of the result, and an opaque candidate reference. It can state a reason for rejection without disclosing internal timeslot or topology details. The response is a snapshot, not a reservation; a later hold may fail if another request consumes a resource. The controller remains responsible for checking actual multiplexing and switching constraints under the applicable OTN technology [G709].

9.3. Outcome and Feedback

The coordinator can record admission outcome, selected path and endpoint identifiers, expected objectives, provisioning state, and observed performance. Feedback permits re-evaluation after faults, load shifts, or compute changes. It should be aggregated to the degree needed for control, limiting disclosure of application and user information.

10. Joint Decision and Control Relationships

Joint coordination relates service objectives, transport feasibility, compute feasibility, and observed outcomes. A realization chooses resources or adapts an existing service within the intersection of authorized domain capabilities. This relationship is applicable to pre-provisioned resources and dynamic provisioning, and does not require a particular reservation sequence.

10.1. Coordination Invariants

Three properties remain stable across realizations. A joint decision preserves the authorized hard constraints of the service. Each resource allocation or change is confirmed by the responsible resource manager. Service assurance accounts for the combined transport and compute outcome. An inference, proposal, or successful internal activity does not establish one of these properties by itself.

Existing shared resources and new allocations have different ownership implications. A candidate can use an existing optical connection or service instance without creating an exclusive allocation. The responsible manager confirms the applicable use and admission conditions. Any reservation, change, or release is limited to the resource portion authorized for that service. Cancelling a service does not authorize tearing down a shared connection or stopping an instance used by other services.

When current information cannot support a required property, the coordination outcome is explicitly uncertain, rejected, or escalated according to policy. Implementations can realize this through different admission, reservation, and recovery mechanisms. The architectural requirement is consistent authority and outcome interpretation, not a universal transaction or state machine.

10.2. Control and Assurance Relationships

Local and domain control may respond to events within their authority and control timescale. Joint control evaluates whether the resulting service continues to satisfy its objective and whether a cross-domain adaptation is warranted. Policy governs stability, permitted disruption, and the balance between an expected benefit and a resource change.

A domain exception, conflicting views, or partially completed resource changes are interpreted against the same authorized service. Resource-state reconciliation remains with the owners, while the JCA maintains the combined service interpretation. Appendix B illustrates one detailed workflow, and Appendix C provides an OTN admission illustration.

11. Agent Cooperation Use Cases

11.1. Interactive Edge Service

A collaboration or inference request has a tight end-to-end delay objective. The closest edge location may be busy, while another location has available compute but a longer optical path. Joint selection compares transport and execution delay rather than minimizing either one in isolation.

The requester identifies an ingress, a service, permitted processing locations, a client rate, and the relevant delay boundary. The operator's profile can supply authorized defaults. When CATS is used, its candidate service contact instances can inform site selection. The coordinator maps eligible contact points to optical transport attachments and obtains service-readiness and capacity assessments from the compute manager. A candidate is excluded when either domain cannot support the service constraints. A new placement or reservation is requested only when needed by the deployment.

For an interactive service with a short activation deadline, the time needed to establish an optical connection may dominate the decision even when the eventual transport delay is low. A pre-established connection can be preferable to a new path that scores better after activation but arrives too late. Once active, small changes in telemetry alone do not justify moving the service, because a site change can disrupt session state or service contact affinity. The service policy determines whether an interruption, migration, or a make-before-break transition is acceptable.

In the cooperative realization, the JCA assigns the latency and activation budget to the TCA and CRA. NEA evidence can expose an unavailable local adaptation even when the abstract route delay is attractive. The CRA can report that a model instance is not yet ready at the nearest site. These results constrain the same path-and-site pair before the JCA chooses a plan, rather than producing unrelated transport and compute optimizations.

11.2. Bulk Data Processing

A data synchronization job has substantial bandwidth demand and a looser delay objective. The coordinator can compare feasible optical connections and compute locations against bandwidth and placement constraints while avoiding unnecessary contention with interactive services.

Here the request can carry an earliest start time, a latest completion time, expected duration or data volume, and a locality restriction. A scheduled admission window may allow the optical transport owner to plan a connection before the workload begins. The coordinator checks that the selected compute site has space for the input data and the processing job, and that the transport owner can sustain the admitted rate for the requested duration. Estimated unused aggregate bandwidth does not substitute for a technology-specific feasibility check for the requested connection.

When several windows or sites are feasible, the ranking policy can prefer lower resource cost or less contention after the hard completion, capacity, and locality conditions have passed. If the job finishes early, the owners release the allocations or service-specific resource shares associated with that job. An overrun requires a new admission decision for any extension; it cannot silently retain capacity beyond the authorized interval.

11.3. Failure or Capacity Change

After an optical network fault or a compute-capacity reduction, the coordinator can consider another feasible pair. Domain controllers still apply their own protection and recovery policies. The coordinator must avoid repeated disruptive changes when transient telemetry fluctuates.

A transport protection switch can restore transport locally while keeping the same compute site. The coordinator first learns whether the service is still meeting its objective and reconciles the active connection reference with the optical transport owner. It does not start a new placement solely because a protected path changed internally. If the remaining connection no longer meets the service contract, or if the compute site fails, it can request a fresh pair using current domain state.

Preparation of a replacement pair may require temporary duplication of transport and compute resources. The coordinator checks whether both owners permit that overlap and whether the service allows state transfer. If the replacement fails after one domain has allocated resources, it releases the new partial allocation and retains the original pair when still usable. If neither the original nor a replacement can meet hard constraints, the outcome is a reported service failure or a degraded mode explicitly authorized by policy, never an unreported relaxation of the original request.

For example, an NEA detects a change in the usable cross-connect resources and reports its affected connection reference and observation time. The TCA verifies the device state and protection outcome, then sends the JCA a domain exception or an updated feasible transport option. In parallel, the CRA reports whether the current workload and alternative placements remain ready. The JCA uses both results to retain the site, move the placement, or report failure. A local diagnosis alone does not trigger a cross-domain resource change.

11.4. Learning-Assisted Policy Adaptation

A control agent may use a DQN as one tool for adapting its local policy to a changing workload mix. The agent obtains authorized observations, passes a declared state representation to the policy, checks the returned action, applies authorized parameter changes through existing controller modules, and reviews the resulting service outcome. Reinforcement learning supplies the decision method inside this use case; the cooperative roles and task boundaries remain applicable to other methods.

The proposed action is checked against current service constraints and resource-owner authority before it can affect resources. The deployment defines the action scope, inference validity, observation freshness, and fallback behavior. Outcome review distinguishes a change in the policy's objective score from a change in the observed service result. The packet-level prototype and its limitations are described in Appendix A.

12. Deployment Considerations

Optical connection setup and workload placement may be much slower than packet forwarding. A coordinator should not assume that each short-lived flow can trigger a new optical connection. Operators can use pre-established connections, admission windows, or batching when setup time exceeds the useful life of a request. The service intent's start time and expected duration help distinguish a candidate for on-demand provisioning from one that needs existing capacity.

In a single transport domain, one optical domain controller may provide the complete feasibility decision. Across domains, an MDSC or comparable coordinator can use abstract topology for candidate ranking, but each domain still confirms its portion of the connection. Domain confidentiality can limit visibility into optical resources; the coordinator should accept an explicit feasible or infeasible result without requiring internal topology disclosure. Administrative policies determine which controller can hold or activate each resource.

Transport and compute metrics evolve on different timescales. A compute load sample and an optical resource snapshot should carry observation time and validity information. The coordinator should define how to handle stale, missing, or contradictory data, including deterministic fallback and rejection when a hard constraint cannot be checked. Re-optimization should account for the delay and disruption cost of changing both an optical connection and a service site.

A minimal deployment can host the JCA, TCA, and CRA as modules in existing coordination and resource-management systems, with NEAs adjacent to selected optical network elements. Roles can be introduced incrementally: legacy elements continue to expose telemetry and configuration through the PNC, while an NEA adds local assessment where useful. A deployment documents which task results depend on agent analysis and which are provided by existing deterministic functions.

For multiple transport domains, each TCA remains responsible for its own NEAs. The JCA receives abstract evidence and candidate references rather than unrestricted element-level context. Failover or agent migration preserves task identity and confirmed owner state. An agent implementation or model update is versioned and evaluated against the same delegation and fallback policy before it is used for active services.

13. Evaluation of Agent Cooperation

Evaluation examines whether a realization preserves the architectural relationships and coordination invariants. Comparable service requests and resource conditions can be evaluated with different abstraction, cooperation, and control placements. Measures include whether local constraints remain visible in domain feasibility, joint decisions preserve authorized objectives, and confirmed resource states agree with service assurance. Improved model reward alone does not demonstrate those properties.

Table 5: Observable cooperation outcomes
Scenario Outcome to observe
Local constraint hidden by abstraction Rejected infeasible pair and retained evidence
Stale or contradictory NEA reports Fresh owner check; no stale-plan activation
Concurrent resource allocation Failed hold and release of the other resource within a specified time limit
Agent restart or delayed result No duplicate write; task recovery from owner state
Local protection restores the path No unnecessary compute move; reconciled service state
Policy tool unavailable or invalid Defined fallback or an explicit failure response

Useful measures include service admission and objective-violation rates, setup and restoration time, task completion and timeout rates, message and tool-call counts, cleanup completion time, stranded resources, and disruptive changes. Comparisons account for the additional observation and delegation cost and report results by topology, load, and failure scenario. Baselines can include existing controller automation without local agents, element agents with only local decisions, and the cooperative architecture under the same resource constraints.

The cooperative architecture has not yet been evaluated with optical transport and computing resources. The existing packet-level results compare policy implementations only and cannot establish the value of agent coordination. These results are recorded in Appendix A.

14. Security Considerations

A forged or exaggerated service intent could consume scarce transport or compute resources. A requester must be authenticated by the deployment's existing mechanisms, and service identifiers, priority, locations, requested capacity, and protection level need authorization before admission. A claimed semantic class cannot confer privileges. Rate limits and quotas can contain repeated expensive feasibility checks or reservation attempts.

False or stale telemetry can lead to incorrect admission, denial of service, or steering to an unintended site. Resource owners need integrity protection for control exchanges, source identification, timestamps, and bounded validity. The coordinator should not treat an unconfirmed model prediction as a substitute for a domain controller's resource check. Learned ranking components require the same feasibility boundary and a deterministic fallback if the model is unavailable or produces an invalid result.

Partial allocation is also a security and availability concern: a failed workload placement must not leave an optical connection indefinitely reserved, and a transport failure must not strand a workload. Holds need expiry, operations need correlated audit records, and recovery needs reconciliation against authoritative domain state. Cross-domain topology abstractions should reveal only the detail required for an authorized decision. This document does not define the security mechanisms for any new interface.

Where agent discovery or task delegation is used, the deployment authenticates the participating identities and binds permitted actions to an owner, resource scope, operation set, and validity interval. A compromised NEA can otherwise supply false local evidence or request inappropriate actions. Controller checks, resource revisions, replay handling, bounded deadlines, and independent confirmation of state-changing outcomes reduce reliance on any single agent report. Task acknowledgements cannot be reused as reservation credentials.

Device descriptions, alarm text, retrieved documents, and peer-agent messages can contain untrusted content. A model-assisted planner treats this content as task data rather than authorization or executable instructions. Tool gateways validate structured arguments against the owner's policy before a write. The deployment constrains delegation depth, concurrent tasks, and tool-call rates to prevent repeated replanning from exhausting control or resource capacity.

Task memory and evidence records are protected from modification and inappropriate disclosure. Model and tool versions, planned actions, owner approvals, actual operations, and observed outcomes are auditable. A model rollback or operator takeover restores a known task owner and checks existing allocations before continuing the service lifecycle.

15. Privacy Considerations

Service classes, attachment points, placement locations, and performance traces can reveal application behavior or user location. Deployments should minimize collected detail, limit retention, and restrict access to the service and operational purpose. Aggregated metrics may be sufficient for candidate ranking. Payload inspection is not required by this architecture, and a service intent should avoid carrying user content or individual traffic records.

Agent context exchange and retained memory may combine application, location, and device information. Each role receives only the evidence needed for its task. Cross-domain results can use opaque resource references and aggregated observations; prompts, traces, and task histories follow the same access and retention policy as their source data.

16. Manageability Considerations

Operators need visibility into rejected constraints, selected path-and-endpoint identifiers, resource hold and activation status, telemetry age, objective violations, cleanup results, and changes initiated by re-optimization. A common request identifier permits correlation across transport and compute logs without exposing internal topology to every participant. Operational metrics can include admission success, setup duration, rollback success, stranded reservation count, and service disruptions. Operators should be able to cap the frequency of disruptive changes and fall back to a deterministic policy if a learned ranking component is unavailable.

A task trace links service intent, parent and child task identifiers, agent identities, plan and policy versions, evidence timestamps, tool operations, owner tokens, and cleanup outcomes. Operators can suspend an agent, revoke delegated write capability, inspect an uncertain operation, or resume through the existing controller workflow. Planned task completion, owner-confirmed resource state, and observed service performance are shown as distinct operational facts.

17. IANA Considerations

This document has no IANA actions.

18. Informative References

[RFC8453]
Ceccarelli, D., Ed. and Y. Lee, Ed., "Framework for Abstraction and Control of TE Networks (ACTN)", RFC 8453, DOI 10.17487/RFC8453, , <https://www.rfc-editor.org/info/rfc8453>.
[RFC8454]
Lee, Y., Belotti, S., Dhody, D., Ceccarelli, D., and B. Yoon, "Information Model for Abstraction and Control of TE Networks (ACTN)", RFC 8454, DOI 10.17487/RFC8454, , <https://www.rfc-editor.org/info/rfc8454>.
[RFC8795]
Liu, X., Bryskin, I., Beeram, V., Saad, T., Shah, H., and O. Gonzalez de Dios, "YANG Data Model for Traffic Engineering (TE) Topologies", RFC 8795, DOI 10.17487/RFC8795, , <https://www.rfc-editor.org/info/rfc8795>.
[RFC10053]
Li, C., Ed., Du, Z., Boucadair, M., Ed., Contreras, L. M., and J. Drake, "A Framework for Computing-Aware Traffic Steering (CATS)", RFC 10053, DOI 10.17487/RFC10053, , <https://www.rfc-editor.org/info/rfc10053>.
[I-D.zhao-cats-otn-applicability]
Zhao, Y., Han, L., Li, X., Zheng, H., and D. King, "Framework and Applicability of Computation-aware Traffic Steering (CATS) in Optical Transport Networks (OTN)", Work in Progress, Internet-Draft, draft-zhao-cats-otn-applicability-03, , <https://datatracker.ietf.org/doc/draft-zhao-cats-otn-applicability/>.
[I-D.li-cats-aisemantic-contract]
Li, Q., Gao, T., and Y. Jiang, "Semantic-Driven Traffic Shaping Contract for AI Networks", Work in Progress, Internet-Draft, draft-li-cats-aisemantic-contract-01, , <https://datatracker.ietf.org/doc/draft-li-cats-aisemantic-contract/>.
[I-D.ietf-ccamp-otn-topo-yang]
Zheng, H., Busi, I., Liu, X., Belotti, S., and O. Gonzalez de Dios, "A YANG Data Model for Optical Transport Network Topology", Work in Progress, Internet-Draft, draft-ietf-ccamp-otn-topo-yang-21, , <https://datatracker.ietf.org/doc/draft-ietf-ccamp-otn-topo-yang/>.
[G709]
ITU-T, "Interfaces for the optical transport network", Recommendation ITU-T G.709/Y.1331, , <https://www.itu.int/rec/T-REC-G.709-202006-I/en>.
[I-D.zhao-ccamp-actn-optical-network-agent]
Zhao, X., Yu, H., Li, A., and Y. Xu, "Integration of Network Management Agent (NMA) into ACTN-Based Optical Network", Work in Progress, Internet-Draft, draft-zhao-ccamp-actn-optical-network-agent-02, , <https://datatracker.ietf.org/doc/html/draft-zhao-ccamp-actn-optical-network-agent-02>.
[I-D.tan-ccamp-onco-control-framework]
Tan, Y., Han, Z., Li, X., Yang, S., and X. Zhao, "A Control Framework for Optical Networks and AI Computing Orchestration (ONCO)", Work in Progress, Internet-Draft, draft-tan-ccamp-onco-control-framework-00, , <https://datatracker.ietf.org/doc/html/draft-tan-ccamp-onco-control-framework-00>.
[I-D.zhao-nmop-network-management-agent]
Zhao, X., Wang, M., Wu, B., Ceccarelli, D., Zheng, H., and J. Zhou, "AI based Network Management Agent(NMA): Concepts and Architecture", Work in Progress, Internet-Draft, draft-zhao-nmop-network-management-agent-05, , <https://datatracker.ietf.org/doc/html/draft-zhao-nmop-network-management-agent-05>.
[RFC6163]
Lee, Y., Ed., Bernstein, G., Ed., and W. Imajuku, "Framework for GMPLS and Path Computation Element (PCE) Control of Wavelength Switched Optical Networks (WSONs)", RFC 6163, DOI 10.17487/RFC6163, , <https://www.rfc-editor.org/info/rfc6163>.

Appendix A. Packet-Level DQN Prototype and Its Limits

This appendix records an existing algorithm prototype as supporting material for the learning-assisted use case. The prototype is a centralized policy adapter using Double DQN updates. It does not implement the proposed control-to-element cooperation, task context, or domain reservation workflow. No reported number is a measurement of agent-cooperation performance.

In the existing packet-level prototype, a centralized DQN observes eleven aggregate features related to link and server queues, server load, delay, recent loss, traffic-class proportions, and topology. It selects one of six predefined control profiles. A profile adjusts service bandwidth shares, routing weights, or the server-load weight; existing heuristics then choose paths and servers. The prototype therefore exercises a control-profile policy; it does not implement direct DQN selection of an optical-connection-and-site pair.

A local, packet-level simulation compared a rule-based route policy, joint route-and-server heuristics, a fixed control action selected on development seeds, and a trained, frozen DQN. With 200 users, two graph topologies, two traffic-shift orders, and three test seeds, the latest evaluation produced mean user satisfaction scores of 83.808, 90.231, 90.644, and 90.564, respectively. The scoring function is specific to that simulation; the values are points on a 0-to-100 scale. The frozen DQN exceeded the topology-tuned fixed policy in four of twelve paired scenarios and did not establish a general advantage.

The simulator models packet links, queues, and servers. It does not model OTN signal structures, optical resource allocation, cross-connect constraints, protection behavior, or provisioning delay. Therefore these numbers are illustrative motivation for joint selection and are not evidence of optical transport performance or optical-network and compute agent cooperation. Technology-specific optical evaluation remains future work.

Appendix B. Illustrative Realization of the Coordination Architecture

This appendix illustrates a task-based realization of the architectural relationships using OTN connectivity and workload placement. The OTN-specific checks apply to this realization, not to every optical switching technology. Its identifiers, holds, activation sequence, and local state names are implementation examples, not mandatory architecture interfaces. Other realizations can use pre-established resources, continuous exchange, or a different orchestration workflow while preserving the properties in Section 10.1.

B.1. Illustrative Task Context and Recovery

The following interactions describe information and responsibility boundaries. They are not new on-the-wire interfaces. A deployment can carry them through existing controller northbound and southbound interfaces, internal APIs, or another authenticated capability-invocation mechanism. An independently deployed agent requires an explicit binding to the responsible owner.

B.1.1. Task Delegation and Result Aggregation

Requester to JCA
The requester supplies authorized service objectives and receives the proposed plan, admission outcome, and service status.
JCA to TCA and CRA
The JCA delegates transport and compute subtasks with the same service correlation and bounded objectives. Each role returns feasibility evidence, a candidate reference, and exceptions; reservation and activation results come from the corresponding owner.
TCA to NEA
The TCA assigns a local assessment or authorized action and receives evidence and observed outcomes. The exchange is bound to the PNC's ownership and device-management context.
Agent to owner tools
An agent invokes existing read or write capabilities according to its scope. The tool outcome is recorded separately from the agent's interpretation of that outcome.

A response distinguishes accepted task, completed assessment, rejected constraint, owner-confirmed hold, confirmed activation, and uncertain execution. These outcomes have different effects on the service lifecycle. An assessment can become stale before a hold, and a successful tool invocation does not prove the end-to-end objective unless the relevant observation confirms it.

B.1.2. Conceptual Information for Cooperation

Table 6: Illustrative task and evidence context
Information Purpose Example use
Service, task, parent task, attempt Correlate decomposition and retries NEA assessment linked to one transport task
Agent and resource-owner identity Identify source and authority PNC responsible for element X
Objective and hard bounds Preserve the authorized intent Capacity, protection, delay, activation deadline
Resource and operation scope Constrain tool use and delegation Assess a cross-connect; no permission to change it
Plan and policy revision Reject obsolete proposals Result for plan revision 2 cannot commit revision 3
Evidence source, time, validity Judge relevance and freshness Resource snapshot and configuration revision
Candidate and owner token Separate assessment from reservation Opaque path reference; time-limited reservation token
Outcome and constraint reason Support aggregation and recovery Local adaptation rejected; activation uncertain
Execution and cleanup reference Reconcile changes in resource state Owner operation identifier and release result

The information is deliberately independent of any agent framework. It can be mapped to existing task and controller records where they already provide the required semantics. An opaque token remains bound to its owner and task; receiving it does not permit an unrelated agent to operate on the resource. Model confidence may accompany an analysis, but it does not substitute for an observation timestamp, resource revision, or feasibility check.

B.1.3. Agent Failure and Task Ownership

One owner is responsible for each task's terminal outcome. If an agent restarts, moves, or becomes unavailable, the replacement recovers task context and queries the resource owner before repeating a state-changing action. Outstanding holds retain their owner-defined expiry. An execution epoch or equivalent owner mechanism prevents a superseded instance from continuing to write resources after task ownership changes.

Missing NEA analysis can fall back to existing device telemetry and PNC checks when they are sufficient for the hard constraints. If required evidence is unavailable, the task fails or is escalated. A failed agent does not justify reporting an unconfirmed allocation as active. Cancellation follows the same resource reconciliation and cleanup procedure as an execution failure.

B.2. Illustrative Provisioning and Assurance Workflow

  1. The JCA authenticates the request and records bounded transport and compute objectives, policy, task correlation, and activation deadline.
  2. The JCA forms candidate OTN-path-and-site pairs and delegates assessments with a defined scope to the TCA and CRA. Both subtasks preserve the same service intent and plan revision.
  3. The TCA gathers the necessary NEA evidence and invokes domain path, adaptation, capacity, switching, and protection checks. The CRA assesses eligible placement and compute readiness through its resource manager.
  4. The JCA aggregates current results. It discards a pair if either domain rejects it or required evidence is missing, then ranks only the remaining feasible pairs.
  5. The selected plan obtains bounded holds from both resource owners. A changed local observation or a failed hold invalidates the affected assessment and triggers cleanup or a fresh plan.
  6. Both owners activate the selected resources through existing control functions. The JCA reports service availability only after both confirmations and any required service-level observation.
  7. NEAs, the TCA, and the CRA report relevant operational changes and action outcomes. The JCA reviews the combined service objective, distinguishing local recovery from a need for joint replanning.
  8. Completion, cancellation, or failure releases only the allocations or resource portions associated with this service and reconciles uncertain outcomes. An explicit exception is reported when cleanup or the service objective cannot be confirmed.

Planning and optimization are subordinate to the resource-owner checks. A control agent can propose an alternative, request more evidence, retain the current service, or return an exception. The resulting operation remains subject to its delegated scope and the current transport and compute state. A learning-based implementation is illustrated in Section 11.4; its algorithm is a local implementation choice.

For an admitted pair, the coordinator retains the normalized hard constraints, the identity and freshness of both domain views, and the reasons that the pair passed each check. Ranking uses only feasible pairs. An apparent improvement in a soft score cannot override a failed adaptation, protection, locality, capacity, or activation-time check. If no pair survives, the result is an explicit rejection with the constraint category that prevented admission; a later retry is a new check against current domain state.

B.2.1. Pair Ranking and Policy Boundary

After admission checks, an implementation can rank feasible pairs by a local objective that combines estimated service delay, resource consumption, setup time, and disruption cost. These quantities have different units and confidence levels; the operator defines their normalization and weights. A lower predicted delay does not necessarily justify a new connection if it consumes scarce protected OTN capacity or requires a disruptive workload move. For an active service, the benefit of a proposed change is compared with the cost of preparation, transition, and possible rollback.

A ranking or adaptation tool can use a fixed rule, a joint heuristic, a learned policy, or a combination. Its declared output determines the next operation: a direct candidate reference selects a pair, while a control-profile action changes scoring or scheduling parameters within configured limits before the existing pair-selection functions run. These action types are not interchangeable. The JCA records which tool and action semantics were used, applies the same deterministic feasibility and authorization checks, and uses a defined fallback or rejects an invalid output. Training reward does not replace resource-owner confirmation.

For comparisons between ranking methods, each method needs the same request stream, resource state, hard constraints, and accounting of rejected requests. A method that improves an average score by silently dropping difficult services or omitting OTN setup cost is not demonstrating better admission. Useful outcome measures include admitted-service fraction, objective violations, setup time, blocked capacity, number of disruptive changes, and cleanup failures, reported by load level and topology. These measures support operational comparison without making any ranking method mandatory for interoperability.

B.2.2. Coordination Lifecycle and Failure Handling

The following states describe a local coordination view, not a protocol state machine: RECEIVED, ELIGIBLE, HELD, ACTIVE, REOPTIMIZING, RELEASING, and FAILED. An implementation may use different internal states. A request identifier correlates retries and prevents duplicate allocations. Capacity and telemetry used for a decision have explicit freshness limits.

 +-------+         +-------+         +-------+         +-------+
 |  JCA  |         |  TCA  |         |  NEA  |         |  CRA  |
 +---+---+         +---+---+         +---+---+         +---+---+
     |                 |                 |                 |
     +--assess path--->|                 |                 |
     |                 +---local task--->|                 |
     |                 |<---evidence-----+                 |
     |<domain result---+                 |                 |
     +------------------assess placement------------------>|
     |<-------------------site result----------------------+
     |                 |                 |                 |
     | Select a pair using BOTH domain results             |
     |                 |                 |                 |
     +---hold path---->|                 |                 |
     |<--PNC token-----+                 |                 |
     +-------------------hold placement------------------->|
     |<------------------compute token---------------------+
     +-activate path-->|                 |                 |
     |                 +--delegated op-->|                 |
     |                 |<local outcome---+                 |
     |<-path active----+                 |                 |
     +-----------------activate placement----------------->|
     |<-----------------placement active-------------------+
     |                 |                 |                 |
     | ACTIVE after BOTH owners confirm  |                 |
     |                 |                 |                 |
     |                 |<-local event----+                 |
     |<-domain event---+                 |                 |
     |<------------placement / outcome event---------------+
     | Review service: retain, replan, or reconcile        |

 TCA uses PNC checks and authorized device tools.
 CRA uses compute manager checks and tools.
 On timeout: query owners. On failure: release partial holds.
Figure 2: Delegation, local evidence, and joint owner-confirmed execution

A hold is a time-limited reservation with an expiry and a domain-specific token. The coordinator verifies that both holds remain valid before activation. If transport setup fails after workload placement, it asks the compute manager to release the placement; if compute activation fails after transport setup, it asks the OTN controller to release the connection. Timed-out or uncertain operations are reconciled against the authoritative domain state before retry. This is compensating cleanup, not a claim of atomic transactions across domains.

The sequence illustrates new allocations for both domains. If a connection or service instance already exists, its manager can instead confirm the permitted use or allocate a service-specific share. Cleanup then releases only the hold or allocation created for this attempt; it does not remove the shared underlying resource. The coordinator records this distinction in the owner-provided operation context.

The illustrative sequence requests a transport hold and a compute hold before activation. A deployment may order these requests according to resource scarcity, expected setup time, and the cost of retaining the earlier hold. Either ordering requires a finite hold lifetime, a remaining activation budget, and cleanup of every earlier hold if a later step fails. The coordinator uses a stable request identifier and attempt number when repeating an operation, and records each returned domain token. A domain that supports idempotent operations can return the prior result for a duplicate attempt. Where such behavior is unavailable, the coordinator queries authoritative state before repeating an uncertain operation so that a delayed response does not create a second allocation.

Activation is confirmed separately by each owner; success from one owner does not imply success from the other. If a deadline expires during an uncertain activation, the coordinator does not report the service as active. It queries both owners, releases resources that are no longer needed, and records any unresolved allocation for operator reconciliation. Concurrent requests can invalidate a feasibility snapshot, so the hold result is the admission decision. Release is also retried or reconciled until the owner reports that the resource is free or an operator takes over an unresolved case.

Table 7: Examples of partial outcomes and recovery
Observed condition Coordinator response State to verify
Compute hold succeeds; OTN hold fails Release compute hold; refresh candidate state Compute capacity restored
Both holds succeed; one expires before activation Stop activation attempt and release remaining hold No resource remains reserved
One activation succeeds; the other fails Compensate by releasing the active resource Owner confirms release or records unresolved case
Activation response times out Query authoritative state before retry or release Exactly which allocation exists
Replacement pair fails during migration Release replacement; retain original when usable Original service and new resources reconciled

During re-optimization, the coordinator checks whether the new pair can be prepared before the old pair is released. A make-before-break transition is used only when both domains support the needed overlap and the resource cost is acceptable. Otherwise the service follows its existing protection or interruption policy. Hysteresis, minimum dwell time, and a change budget limit churn when telemetry fluctuates. A domain fault can trigger local protection without waiting for the coordinator; later coordination reconciles the resulting path and service state.

NEA and CRA events are correlated with the active service and resource revision. A late result for an abandoned plan cannot authorize its activation. When an NEA reports a local change, the TCA determines whether the existing domain connection remains valid and whether protection already restored it. The JCA replans only when the combined service state and authorized change policy require it.

Appendix C. Illustrative OTN Feasibility Realization

Consider an authorized request for a client signal with at least 5 Gbit/s of admitted transport capacity, no more than 15 ms of combined one-way OTN transport and execution delay, link-disjoint working and protection connections, and activation within 30 seconds. For this example, ingress access and other processing delays have already been budgeted separately by the service contract. The requester is authorized for two edge sites. The compute manager reports one available workload placement at each site. The values below are illustrative estimates and local resource counts, not standardized OTN rates or measured equipment performance.

Table 8: Candidate site comparison
Site Transport Compute Decision
A 6 Gbit/s; 8 ms 10 ms Reject: estimated total 18 ms
B 6 Gbit/s; 6 ms 7 ms Continue to OTN check: estimated total 13 ms

The OTN controller's operator-local planning profile maps this client request to an ODUflex-capable connection that requires four tributary resource units on every segment of each protected route. The number four is an assumption of this example's local planning profile; it is not a G.709 tributary-slot conversion rule. Both ingress and egress must support the selected client adaptation and connection type. The controller checks these interfaces, the switching matrices, the two routes, and their protection relationship before returning a feasibility result.

Table 9: Illustrative OTN resource ledger for site B
Check Domain observation Result
Client adaptation and endpoints Ingress and B egress support the selected adaptation and ODUflex connection Pass
Working route via X Four free units on each segment and through the X cross-connect Pass
Protection route via Y Four free units on each segment and through the Y cross-connect; links distinct from working route Pass
Setup budget Estimated connection setup 20 seconds Pass against 30-second bound
                         working: 4 local units
                    +-----------------------+
             +----->| X working cross-connect|----+
             |      +-----------------------+    |
  +----------+----+                           +---v----------+
  | Ingress P0    |                           | Egress B     |
  | client signal |                           | edge site B  |
  +----------+----+                           +---^----------+
             |      +-----------------------+    |
             +----->| Y protection cross-   |----+
                    |   connect             |
                    +-----------------------+
                       protection: 4 local units

       X and Y use link-disjoint routes in this example.
       Unit counts are local assumptions, not G.709 rates.
Figure 3: Illustrative protected OTN paths to site B

Only after this check does B become an admissible pair. The coordinator holds the workload placement and asks the OTN controller to hold the protected connection, each with an illustrative 60-second expiry. If both holds succeed while the 30-second activation budget still permits completion, it activates both resources and reports the service active after both owners confirm. The hold expiry is a safety limit, not permission to exceed the requested activation deadline.

Suppose a concurrent request consumes two units on a working-route segment after the feasibility response but before the OTN hold. The controller now has only two free units on that segment and rejects the hold. The coordinator releases the compute hold and does not report B as admitted. It may seek a fresh candidate or reject the request; it does not silently remove the requested protection. Similarly, if the X cross-connect had only two usable units at the initial check, the abstract 6 Gbit/s path estimate would not make the connection feasible. This example illustrates the distinction among metric-based ranking, technology-specific OTN feasibility, and a resource hold that survives concurrent allocation.

In the agent realization of this example, the JCA assigns a transport assessment to the TCA and a site-readiness assessment to the CRA. The TCA requests local checks with a defined scope from the NEAs at ingress, X, Y, and egress B. Each NEA returns evidence for the same task and plan revision, including the observed configuration and resource validity. The TCA uses this evidence with the PNC's authoritative connection and protection checks before returning B's domain candidate reference. The CRA obtains placement readiness from the compute manager. The JCA correlates these results before requesting either hold.

If the NEA at X reports a newer configuration with only two usable units, the TCA withdraws or recomputes the earlier candidate even if a stale report from Y still appears favorable. If an activation command later times out, a completed agent subtask does not prove that the cross-connect exists: the TCA queries the PNC and device state and the JCA reconciles the compute allocation. These cases illustrate how local evidence and controller authority participate in the joint decision.

Authors' Addresses

Hui Yang
Beijing University of Posts and Telecommunications
10 Xitucheng Road, Haidian District
Beijing
Beijing, 100876
China
Mingfeng Han
Beijing University of Posts and Telecommunications
10 Xitucheng Road, Haidian District
Beijing
Beijing, 100876
China
Qiuyan Yao
Beijing University of Posts and Telecommunications
10 Xitucheng Road, Haidian District
Beijing
Beijing, 100876
China
Tiankuo Yu
Beijing University of Posts and Telecommunications
10 Xitucheng Road, Haidian District
Beijing
Beijing, 100876
China