<?xml version='1.0' encoding='utf-8'?>
<!DOCTYPE rfc [
  <!ENTITY nbsp    "&#160;">
  <!ENTITY zwsp   "&#8203;">
  <!ENTITY nbhy   "&#8209;">
  <!ENTITY wj     "&#8288;">
]>
<?xml-stylesheet type="text/xsl" href="rfc2629.xslt" ?>
<!-- generated by https://github.com/cabo/kramdown-rfc version 1.7.43 (Ruby 3.4.12) -->
<rfc xmlns:xi="http://www.w3.org/2001/XInclude" ipr="trust200902" docName="draft-wnd-opsawg-icon-ps-02" category="info" consensus="true" submissionType="IETF" tocInclude="true" sortRefs="true" symRefs="true" version="3">
  <!-- xml2rfc v2v3 conversion 3.35.0 -->
  <front>
    <title abbrev="Observability and I&amp;C">Problem Statement for Autonomous Agent Observability, Intervention and Control (I&amp;C) in Multi-Agent Autonomous Networks</title>
    <seriesInfo name="Internet-Draft" value="draft-wnd-opsawg-icon-ps-02"/>
    <author fullname="Qin Wu">
      <organization>Huawei</organization>
      <address>
        <email>bill.wu@huawei.com</email>
      </address>
    </author>
    <author fullname="Daniele Ceccarelli">
      <organization>Cisco</organization>
      <address>
        <email>daniele.ietf@gmail.com</email>
      </address>
    </author>
    <author fullname="Zhenqiang Li">
      <organization>CMCC</organization>
      <address>
        <email>li_zhenqiang@hotmail.com</email>
      </address>
    </author>
    <author fullname="Luis. M. Contreras">
      <organization>Telefonica</organization>
      <address>
        <email>luismiguel.contrerasmurillo@telefonica.com</email>
      </address>
    </author>
    <author fullname="Daniel King">
      <organization>Lancaster University</organization>
      <address>
        <email>d.king@lancaster.ac.uk</email>
      </address>
    </author>
    <date year="2026" month="November" day="16"/>
    <keyword>agent observability</keyword>
    <keyword>ai guardrail</keyword>
    <keyword>agent intervention and control</keyword>
    <abstract>
      <?line 83?>

<t>This document provides an overview of the issues associated with the
deployment of the observability, intervention, and control of autonomous
agent pipelines in large-scale heterogeneous network environments. The
term "Observability, Intervention and Control" is used to describe a set of automated and
human-initiated mechanisms that guarantee the capability to observe, evaluate, constrain,
correct, and terminate Autonomous agents at any point, for any reason, irrespective of
their level of autonomy under which it operates, to ensure resilience, recovery,
and operational continuity.</t>
      <t>The set of enabled observability, intervention and control reflects operator
service offerings to ensure that autonomous network operations can be stopped, or safely
redirected when required and is designed in conjunction with agent to agent, agent
to tools, agent to human interaction and service and network policy.</t>
      <t>This document also identifies several key areas that the Agent Observability,
Intervention and Control group will investigate to guide its
architectural and protocol work and associated documents.</t>
    </abstract>
  </front>
  <middle>
    <?line 103?>

<section anchor="introduction">
      <name>Introduction</name>
      <t>Network operations are increasingly autonomous with the growth of AI Agent
applications at both network level and service level <xref target="IG1507"/>, <xref target="IG1251G"/>.
The Agent lifecycle management comprise the following phases:</t>
      <ul spacing="normal">
        <li>
          <t>Agent Discovery: Discover capabilities and skills and onboard agent</t>
        </li>
        <li>
          <t>Agent Deployment: move agent from pilot project to production environments</t>
        </li>
        <li>
          <t>Agent Upgrade: Large language model, tools, prompts, memory related software update</t>
        </li>
      </ul>
      <t>To help network operators manage Autonomous Agents with more consistency,
visibility and control, the observability component, intervention and control components
need to work in a collaborate manner and are critical to provide guardrail protection for
the Agent lifecycle management.</t>
      <t>Since AI native operations may be non-deterministic, when Autonomous Agents such as network management agents
misbehave or deviate from what Agents are expected to do, current AI control
technologies (often referred to as "AI guardrails") are introduced to constrain
the behavior of Autonomous Agents within operational and compliance boundaries, prevent
AI from producing harmful results or taking wrong actions, e.g., escalate a decision
to a human for a high-risk network operation, defend against malicious attacks,
e.g., prompt injection. These AI guardrails enable you to do checks and validations of user
input and agent output and typically break down into input guardrail, action guardrail,
output guardrail and operate at the input/output/pre-action filter level with static boundary
parameters. For example, image you have an Autonomous Agent that uses network active
and reactive assurance component to identify and resolve network issues, ensuring that any
disruption or degradations are promptly addressed. You wouldn't expect operational anomalies
or performannce drifts of agent behavior produce harmful results, leading to high risk network
operation failure. If the guardrail detects such operational anomalies, it can immediately
raise an error and prevent harmful results of high risk network operators.</t>
      <t>However, as Agentic AI systems are increasingly integrated into autonomous workflows and
critical infrastructure, these static measures are proving insufficient for the
full operational lifecycle, e.g.,</t>
      <ul spacing="normal">
        <li>
          <t>Unable to detect, interrupt, and rollover from unanticipated behaviors;</t>
        </li>
        <li>
          <t>Network operators usually lack an equivalent infrastructure or platform for human oversight;</t>
        </li>
        <li>
          <t>Unable to provide continuous monitoring of an AI system's internal logic or its long-running
execution paths that match the speed and scale of the Autonomous Agent applications,
e.g., network failure or security risk is hard to detect and control, occurring at machine speed.</t>
        </li>
        <li>
          <t>When a violation related to input/output filter is suspected, there are currently no standardized
interoperable protocols for network operation and task intervention (e.g., immediate task suspension)
and recovery (e.g., reverting to a last known safe state or undoing a series of autonomous actions
that introduce substantial operational risk) mechanisms.</t>
        </li>
        <li>
          <t>In non-deterministic environments, the lack of human oversight and human-AI semantic intent exchange
hinder timely risk mitigation and state recovery during boundary violations by agents.</t>
        </li>
      </ul>
      <t>This document provides a problem statement for protocol on continuous agent behavior observability, intervention and control.
We list the properties the protocol should have, then explain why those properties are necessary. We explain why
a new protocol can offer the optimal solution for the more general problem of identifying and characterizing trajectory records
related to agent behavior or workflow operation, continuous monitoring and evaluation, enable human oversight, provide
human and agent interaction for agent intervention and control at the service level and network level.</t>
      <t>Where possible, any solutions work will be built in a modular way using existing IETF protocols.
However, no protocol solution choices will be made until the functional requirements have been
agreed, and then this will require an analysis of the capabilities of existing protocols and
identify gaps that need to be filled.</t>
    </section>
    <section anchor="conventions-and-definitions">
      <name>Conventions and Definitions</name>
      <t>The key words "<bcp14>MUST</bcp14>", "<bcp14>MUST NOT</bcp14>", "<bcp14>REQUIRED</bcp14>", "<bcp14>SHALL</bcp14>", "<bcp14>SHALL
NOT</bcp14>", "<bcp14>SHOULD</bcp14>", "<bcp14>SHOULD NOT</bcp14>", "<bcp14>RECOMMENDED</bcp14>", "<bcp14>NOT RECOMMENDED</bcp14>",
"<bcp14>MAY</bcp14>", and "<bcp14>OPTIONAL</bcp14>" in this document are to be interpreted as
described in BCP 14 <xref target="RFC2119"/> <xref target="RFC8174"/> when, and only when, they
appear in all capitals, as shown here.</t>
      <?line -18?>

<ul spacing="normal">
        <li>
          <t>Autonomous Agent: An AI-driven software entity capable of accepting
 a declarative goal and executing non-deterministic tasks.</t>
        </li>
        <li>
          <t>Agent Drift: Gradual performance degradation or misalignment of
 reasoning patterns over time in a production environment.</t>
        </li>
        <li>
          <t>Cascading Failure: A scenario where a failure in one downstream
 agent propagates across multi-agent boundaries.</t>
        </li>
        <li>
          <t>Human to Agent Communication: The interaction between human users and Autonomous Agent designed to perform tasks, solve
  problems, or provide information. Unlike standard human-to-machine interaction where a human drives every step of a
  task, human-agent communication involves delegation, where the human provides a goal and the agent autonomously figures out
  how to achieve it.</t>
        </li>
        <li>
          <t>Observability: The visibility into an agent's internal state, decision-making logic, and workflow execution from its
 external telemetry outputs (e.g., logs, traces, metrics), enabling human   operators or monitoring systems to understand what
 the agent is doing and why it behaves in a specific manner.</t>
        </li>
        <li>
          <t>Intervention: A reactive, emergency action to intervene or take control of an agent with boundary violations, anomalies, failures,
              or risks, so as to block harmful decisions, disrupt hazards, malicious abuse, and promptly mitigate losses.</t>
        </li>
        <li>
          <t>Control: Establish a deterministic operational boundary for the agent before execution. By pre-defining the agent's behavior scopes,
         operational constraints, and security baselines, it fundamentally mitigates abnormal behaviors from agents.</t>
        </li>
        <li>
          <t>Evaluation: Using Trajectory Record to assess the performance and understand how an Agent solves problems, e.g., checking if the
        agent took the shortest sequence of actions or wasted resources on redundant tools or analyzing specific segments of the
        trajectory to see if the agent excels at information retrieval but struggles with mathematical synthesis.</t>
        </li>
        <li>
          <t>Human Oversight/Human on the Loop: The practice of keeping humans actively involved in continuously monitoring of Autonomous agents.
                 In the agent trajectory management, itt ensures that Autonomous Agents do not go off the rails, violate
                 safety protocols, or waste resources. It transforms a fully autonomous "black box" into a controllable and
                 collaborative system.</t>
        </li>
        <li>
          <t>Behavior: pattern of reasoning, decisions, and actions an Autonomous agent takes to achieve a specific goal such as reasoning sequence,
          the sequence and logic of execution paths.</t>
        </li>
        <li>
          <t>Trajectory Record: Keep track of Agent behaviour and produce audit log or trace information to Capture the entire "flight path"
                   or reasoning sequence the agent followed to reach its conclusion using a structured Thought,Action,Observation
                   loop.</t>
        </li>
        <li>
          <t>OpenTelemetry: also known as OTel, is a vendor-neutral open source Observability framework for instrumenting, generating, collecting,
and exporting telemetry data such as traces, metrics, and logs ({OpenTelemetry}).</t>
        </li>
      </ul>
    </section>
    <section anchor="problem-space">
      <name>Problem Space</name>
      <t>The deployment of autonomous agentic systems within operators' networks introduces fundamental operational,
architectural challenges. Current network management paradigms are static rule driven and therefore are built on
deterministic models that assume predictable, rule-based behaviors. The shift toward non-deterministic (probabilistic),
AI-driven network operation architectures creates a structural mismatch between machine-speed execution and human-speed
oversight. This gap appears in the following three distinct aspects.</t>
      <section anchor="the-observability-aspect">
        <name>The Observability Aspect</name>
        <section anchor="limited-transparency-in-planning-reasoning-and-tools-execution">
          <name>Limited Transparency in Planning, Reasoning, and Tools Execution</name>
          <t>As agents increasingly execute complex operational tasks (e.g., service
provisioning, fault diagnosis), they frequently delegate critical planning
paths and execution decisions to the underlying Large Language Models (LLMs)
or sub-task agents. This delegation creates an optimization barrier:</t>
          <ul spacing="normal">
            <li>
              <t><strong>Lack of Trajectory Transparency:</strong> Network operators cannot cannot
 validate the safety or intent of an agent's planned mutations and trace how
 an agent’s internal Chain-of-Thought (CoT) reasoning maps directly to mutating
 network configuration diffs (e.g., CLI changes or Network configuration changes)
 before or after execution. This may lead to unexpected consequence on the
 infrastructure.</t>
            </li>
            <li>
              <t><strong>Ambiguity of Data Attribution:</strong> There is no standardized mechanism to observe
 where context or knowledge come from. Operators cannot verify which external
 knowledge base, version, or retrieval weight led an agent to make a
 high-risk operational decision.</t>
            </li>
            <li>
              <t><strong>Invisible Tool Parameter Bindings:</strong> Input and output states of invoked tools,
 scripts, and APIs are encapsulated within proprietary agent execution loops,
 preventing real-time validation of parameter bindings and performance evaluation.</t>
            </li>
          </ul>
        </section>
        <section anchor="ambiguity-of-accountability-or-responsibility-determination">
          <name>Ambiguity of Accountability or Responsibility Determination</name>
          <t>In distributed agent interaction (e.g., agent to tools, human to agent, agent
to agent) topologies, operational responsibility for an ultimate network
outcome is scattered across an extended chain of coordinating agents,
foundational models, and abstraction layers. When system failures, performance
degradations, or unintended consequences occur, attributing accountability to
a specific agent entity, localized model decision, or human-in-the-loop anchor
becomes highly ambiguous. This lack of clear log, trace and performance metrics characterizing
agent operational health such as action execution latency, error rates, creates severe
complications for post-incident root-cause analysis and regulatory compliance reporting.</t>
        </section>
        <section anchor="high-velocity-telemetry-data-ingestion-inefficiency-for-evaluation-and-intervening">
          <name>High-Velocity Telemetry Data Ingestion Inefficiency for Evaluation and Intervening</name>
          <t>Agents are explicitly designed to operate with high degrees of autonomy, speed, and
scale. However, network operators currently lack the corresponding telemetry
mechanism and human-on-the-loop infrastructure required to observe, evaluate, and intercept
these systems at the same machine-speed pace.  Consequently, effective
real-time oversight becomes functionally impossible. Always relying on human
escalation paradigms is usually impractical due to the sheer volume and velocity
of the data points involved in active agent pipelines:</t>
          <ul spacing="normal">
            <li>
              <t><strong>Asynchronous Tracking and Clock Drift:</strong> Distributed environments lack an authoritative
 causal event-ordering model. This prevents network operators from cleanly aligning internal agent
 reasoning loops chronologically with external network state telemetry changes.</t>
            </li>
            <li>
              <t><strong>Telemetry Storms:</strong> Intensive CoT reasoning logs and high-frequency tool invocation
 traces can lead to telemetry storms that overwhelm Agent Observability Data collectors, yet legacy pipelines
 lack adaptive backpressure or dynamic sampling mechanisms.</t>
            </li>
            <li>
              <t><strong>Security vs. accountability Trade-off:</strong> Standard telemetry logging lacks context-aware, dynamic
 update. Network Operators are forced to choose between logging complete trajectories (risking
 the leakage of PII, credentials, or topology details) or omitting data (ruining post-incident
 accountability and cryptographic audit trails).</t>
            </li>
          </ul>
        </section>
      </section>
      <section anchor="the-control-aspect">
        <name>The Control Aspect</name>
        <section anchor="inadequacy-of-boundary-constraints-and-risk-evaluation">
          <name>Inadequacy of Boundary Constraints and Risk Evaluation</name>
          <t>Agent behavior cannot be reliably constrained using predefined, deterministic
rules or traditional static guardrails. Because agents rely on dynamic reasoning
patterns to achieve declarative goals, their exact execution trajectories remain
non-deterministic. This intrinsic variability bypasses legacy input/output filters
that fail to account for real-time contextual adaptation and makes operational
outcomes significantly less predictable during runtime execution, leading to
multi-step execution risks:</t>
          <ul spacing="normal">
            <li>
              <t><strong>Implicit Impact Evaluation:</strong> Autonomous Agents lack a standardized
 protocol to evaluate and classify the risk tier (e.g., Low, Medium, High, Critical)
 or estimate the potential blast radius (affected devices, link traffic, customer
 scope) of an action sequence before deploying in production network.</t>
            </li>
            <li>
              <t><strong>Automation Storms:</strong> Autonomous reasoning can lead to high-velocity loops.
  Without standardized per-device, per-operator access control, or global
 concurrency access controls and rate limits, agents run the risk of losing control
 on the network elements.</t>
            </li>
            <li>
              <t><strong>Undefined Action Cancellation Semantics:</strong> Existing control frameworks do not
   natively support state-aware cancellation transitions,leading to partial,
   configurations or applied configuration without monitoring.</t>
            </li>
          </ul>
        </section>
        <section anchor="static-identity-management-and-access-control">
          <name>Static Identity Management and Access Control</name>
          <t>Traditional Identity and Access Management (IAM) frameworks were designed exclusively
for human operators or static, deterministic software processes. These frameworks cannot
securely tackle emerging dynamic agent attributes such as autonomous entity identities
and behavioral profiles which are frequently created and modified at runtime rather than
being pre-provisioned with metadata information to describe functional capabilities.</t>
          <ul spacing="normal">
            <li>
              <t><strong>Absence of Role Separation:</strong> Existing agent frameworks lack the mechanisms to enforce
 strict, credential Separation of Duties across distinct operational roles (Read-Only,
 Propose, Validate, Approve, Execute). This frequently allows a single agent to plan,
 approve, and execute a network configuration changes end-to-end without supervisory
 intervention.</t>
            </li>
            <li>
              <t><strong>Coarse Permission Demarcation:</strong> Security policies cannot dynamically bind/constrain
  an agent's permissions to a specific agent in the network domain based on task context,
  nor can they safely manage downstream sub-agent permission delegation or context-dependent
  privilege escalation.</t>
            </li>
            <li>
              <t><strong>Concurrency and Policy Conflicts:</strong> In federated multi-vendor topologies, multi-agent
  operations lack distributed resource locking. When an action spans multiple network
  domains (e.g., routing vs. security), there is no deterministic framework to resolve
  cross-domain policy conflicts or enforce authority precedence (e.g., ensuring a Human
  Operator or Supervisor Agent explicitly preempts a lower-level autonomous agent).</t>
            </li>
            <li>
              <t><strong>Control Unreachability Vulnerability:</strong> There is no standardized "fail-safe" or
  "fail-hold" expectation when an agent loses connectivity to its management plane,
  meaning a disconnected agent may continue altering network state entirely unconstrained.</t>
            </li>
          </ul>
        </section>
        <section anchor="fragmentation-across-heterogeneous-integration-layers">
          <name>Fragmentation Across Heterogeneous Integration Layers</name>
          <t>Agentic systems operating across mixed Operational Support Systems (OSS)
and Network Management domains must interact with a highly heterogeneous
mix of legacy systems, modern APIs, and third-party platforms. Establishing
consistent, operational and compliance boundaries across these disparate
integration layers is exceptionally complex as agents may routinely validate intent,
invoke actions or retrieve data through pathways, e.g. <xref target="MCP"/> and <xref target="A2A"/>,
that were never designed with network management automation we used today.</t>
        </section>
        <section anchor="multi-vendor-dependency-risks">
          <name>Multi-Vendor Dependency Risks</name>
          <t>As network operators begin sourcing agentic control capabilities from diverse third-party
vendors, independent software providers, and hyperscalers, operational
accountability becomes externalized in ways that are difficult to technically
or contractually enforce. A typical production agent implementation features
highly fragmented dependency chains spanning completely separate vendor
ecosystems, including LLM/AI model providers, infrastructure hosts, core
agent frameworks, tool/API repositories, and interconnection fabrics.</t>
        </section>
      </section>
      <section anchor="the-intervention-aspect">
        <name>The Intervention Aspect</name>
        <section anchor="lack-of-human-oversight-capabilities-such-as-rollback-and-recover">
          <name>Lack of Human Oversight Capabilities such as Rollback and Recover</name>
          <t>When an autonomous agent deviates from expected boundaries (e.g., entering
infinite reasoning loops, suffering from reasoning drift, or getting stuck
in deadlocks), the current human oversight capabilities including runtime
intervention for execution interruption, deterministic transaction rollback
and Recovery, and human escalation remain highly immature and lack
interoperability:</t>
          <ul spacing="normal">
            <li>
              <t><strong>Lack of Out-of-Band Kill Switches:</strong> The legacy Frameworks lack an
 asynchronous, out-of-band communication interface that can force an
 uncooperative or frozen agent execution loop to immediately pause or
 stop, completely independent of its internal LLM responsiveness.</t>
            </li>
            <li>
              <t><strong>Information Loss During Escalation:</strong> When an agent encounters an
 ambiguous state and needs to escalate a problem to a higher human
 authority, it lacks an interoperable protocol to encapsulate and
 pass its full reasoning provenance and context trail seamlessly.</t>
            </li>
            <li>
              <t><strong>Lack of Granularity Control in Recovery:</strong> Operators lack standard
 mechanisms to trigger targeted rollbacks. They cannot choose
 whether to revert a single execution step (Workflow level),
 a distinct operation (Task level), or an entire multi-step
 execution (Context level).</t>
            </li>
            <li>
              <t><strong>Dangerous Infrastructure Overrides:</strong> Operators must rely on
 primitive infrastructure-level actions (e.g., killing a process,
 revoking an API key, or suspending a service account). These
 actions wipe out the runtime memory, provide no out-of-band
 state preservation, and allow no selective pausing or soft
 redirection.</t>
            </li>
          </ul>
          <t>Without human oversight capabilities, the existing network operation
can not ensure any disruptions or degradations are promptly addressed
which lead to harmful outcome for high risk network configuration
changes.</t>
        </section>
      </section>
    </section>
    <section anchor="solution-space-for-autonomous-agent-observability-intervention-and-control">
      <name>Solution Space for Autonomous Agent Observability, Intervention and Control</name>
      <section anchor="opentelemetry-for-agent-observability-and-drift-tracking">
        <name>OpenTelemetry for Agent Observability and Drift Tracking</name>
        <t>Modern agents orchestrate complex workflows: reasoning chains, tool execution, knowledge
retrieval, multi-agent collaboration. When things go wrong, or right, you need to understand
exactly what happened. Traditional network monitoring such as gRPC, SNMP, YANG Push, IPFIX
can't capture reasoning processes or decision context. OpenTelemetry {OpenTelemetry} addresses
this by utilizing unified GenAI and Agent Semantic Conventions to standardise how metrics, logs,
and distributed traces are captured across multi-agent system.</t>
        <t>Implementing OpenTelemetry for Autonomous Agents focuses heavily on distributed tracing and Context
Monitoring to how an agent processes information, arrives at decisions, and executes tasks
to establish baselines and identify drift early as follows:</t>
        <ul spacing="normal">
          <li>
            <t>Distributed Tracing (Spans): The agent as a whole run acts as the root span. Every individual
reasoning loop, agent delegation, LLM invocation, and tool/API execution is mapped as a child span.
This layout instantly reveals where latencies, bottlenecks, or errors occur.</t>
          </li>
          <li>
            <t>GenAI Semantic Conventions: Standardised metadata tags provide explicit context. Spans automatically
record critical variables across four critical domains: System Context, Token Economics, Vector Retrieval,
and Agent Reasoning,like gen_ai.request.model, gen_ai.usage.input_tokens, and gen_ai.usage.output_tokens.</t>
          </li>
          <li>
            <t>Protocol, Decision and System Events: When opted-in, Opentelemetry logs every agent actions, every decision,
every protocol communication between agents or between agent and tools. This visibility allows engineers to
review the exact context that caused an agent to exhibit non-deterministic behavior or get stuck in an infinite
loop.</t>
          </li>
        </ul>
      </section>
      <section anchor="context-aware-ai-guardrails-and-validation">
        <name>Context-Aware AI Guardrails and Validation</name>
        <t>These are most mature, and most operationally familiar AI control mechanism in production today. AI Guardrail
approaches currently realized in the industry operate at defined transition points in the agent pipeline, primarily
prompt filtering at the LLM input boundary, response validation at the LLM output boundary, and access control
restrictions on tool invocation boundary. Currently, AI Guardrails are checks that run alongside your agents to catch
bad input or bad output — without necessarily involving your selected large language model(expensive or cheap).
For example, image you have an Autonomous Agent that uses network active
and reactive assurance component to identify and resolve network issues, ensuring that any
disruption or degradations are promptly addressed. You wouldn't expect operational anomalies
or performannce drifts produce harmful results, leading to high risk network operation failure.
if the guardrail detects such operational anomalies, it can immediately raise an error and
prevent harmful results of high risk network operators.</t>
        <t>AI Guardrail are realized through 4 different mechanisms:</t>
        <ul spacing="normal">
          <li>
            <t>Rule based filters: Apply pattern matching, keyword blocking, regular expressions, and deterministic logic to
prompts, context retrieval and completions.</t>
          </li>
          <li>
            <t>LLM based safety classifiers: Use a secondary language model to evaluate the primary model's output for safety
policy compliance before it is returned.</t>
          </li>
          <li>
            <t>Agent framework based guardrail libraries: Provide structured policy specification languages (e.g. NeMo Guardrails)
that allow developers to express policy rules in a higher-level format, with the framework handling enforcement
logic.</t>
          </li>
          <li>
            <t>Prompt engineering constraints: shape model behaviour by instruction rather than by interception.</t>
          </li>
        </ul>
        <t>AI Guardrails enforce policy constraints at boundaries to prevent non-deterministic or harmful network operations.
Take Multi-Step &amp; Hybrid Defense as an example, AI Guardrail Combines rule-based filters, LLM safety classifiers,
and frameworks like NeMo Guardrails to evaluate multi-agent alignment across pipelines.</t>
      </section>
      <section anchor="agent-drift-detection">
        <name>Agent Drift Detection</name>
        <t>Agent drift refers to the gradual degradation in an agent's performance and alignment with intended behavior over time
in production environments. It exists in many forms depending on the aspect of the system affected. For example, goal
drift occurs when the quality of goal execution deviates from expectations, context drift when the relevance or
accuracy of the working context deteriorates; reasoning drift when there is a decline in the agent's planning and
decision-making capability, and collaboration drift when the effectiveness of interactions with tools, external APIs,
or other agents degrades.</t>
        <t>Agent drift is a well-recognized problem in academic research and agent frameworks. However, there is no universally
applicable control mechanism that addresses all scenarios in practice. A key reason for this is the
strong dependence on domain-specific expertise and observability mechanisms to detect, diagnose, and mitigate drift
effectively. Many of these also may require fine-tuning the base model with revised data sets. So a runtime control of
drift needs to be addressed in a case-by-case basis. Some of the practices followed for addressing the Agent drift are
as follows</t>
        <ul spacing="normal">
          <li>
            <t>Goal drift is observed by statistical evaluation of production tasks against the evaluation tasks. The mitigation
may involve fine-tuning the model with revised task lists and associated agent performance.</t>
          </li>
          <li>
            <t>Context drift is detected by monitoring the retrieved context, retrieval parameters/metrics and mitigated through
context refresh, context window management and memory management.</t>
          </li>
          <li>
            <t>Reasoning drift is detected through metrics such as relevance, success rate, tool selection/usage, LLM-as-a-judge
and it is mitigated through fine-tuning of models, optimizing the prompts/prompt engineering</t>
          </li>
          <li>
            <t>Collaboration drift is typically detected through interaction success rates across agent interactions and mitigated
by fixing the issues with tool/API/agent interactions.</t>
          </li>
        </ul>
      </section>
      <section anchor="quality-gates-for-human-oversight-stage-transitions">
        <name>Quality Gates for Human Oversight &amp; Stage Transitions</name>
        <t>Quality gates are checkpoints that evaluate whether an operation should proceed or not, or should be conditionally
allowed.</t>
        <t>Unlike Guardrails which enforce policy constraints at defined boundaries of Agent implementation, Quality
Gates assess whether the work product of one stage meets a defined quality standard before permitting progression to
the next. The concept is borrowed from DevOps practice i.e. quality gates in CI/CD pipelines that prevent code from
advancing through build, test, and deployment stages unless it meets defined quality criteria. While guardrails determine
crossing points i.e. what enters and exits defined zones, quality gates determine progression points - whether work of
sufficient quality advances to the next stage.</t>
        <t>Quality gates are the ideal mechanism to involve humans for agent tasks
execution quality asessment, i.e., at stage transitions where the accumulated work product of a whole reasoning
stage is ready for assessment where the human is presented with a complete plan, a complete risk assessment, and a
specific decision to make.</t>
        <t>Currently none of the available agent frameworks have a named capability called quality gate. However, some of the
existing functionality can be leveraged for realizing this. For example LangGraph has concept of conditional edges that
enable dynamic, non-linear workflows which allows routing execution to different nodes based on a state-evaluating
function or if-else statements. Similarly, Google ADK (Agent Development Kit) provides callback that can be invoked
before or after tool use and implement quality gate logic. So most of the techniques that exist today are agent
framework specific.</t>
        <section anchor="existing-intervention-approaches-for-human-and-agent-interaction">
          <name>Existing Intervention Approaches for Human and Agent Interaction</name>
          <t>Modern network agents orchestrate complex workflows that occasionally run into infinite reasoning loops, reasoning drifts,
or operational deadlocks. When boundaries are violated or an agent begins to misbehave, operators require a clear mechanism
to step in and pause, correct, or safely redirect the agent. Traditional infrastructure-level overrides (such as process
kills or API credential revocations) cannot gracefully manage context state, leaving network configurations partially applied.
Existing intervention mechanisms address this by establishing structured Human-in-the-Loop (HITL) communication models to
bridge the gap between autonomous task execution and external human oversight.</t>
          <t>Implementing intervention approaches for Autonomous Agents focuses heavily on task suspension, state-aware recovery, and runtime
command injection to intercept failures early through the following mechanisms:</t>
          <ul spacing="normal">
            <li>
              <t>Infrastructure-Level Controls (Repurposed): The most immediate, though primitive, intervention approach borrows
standard IT infrastructure management actions. Operators forcefully apply process termination, API key revocation,
or service account suspension to stop an agent. While highly effective at freezing uncooperative pipelines, these
brute-force approaches wipe out the agent’s runtime memory and provide no mechanism for graceful recovery or
selective sub-task modification.</t>
            </li>
            <li>
              <t>Framework-Specific Interaction Primitives: Modern orchestration frameworks provide embedded, code-level capabilities
to pause and resume agent execution streams. For example, frameworks like LangGraph utilize native checkpointing states
to let an agent pause its reasoning cycle, seek guidance or explicit input from a human, and subsequently proceed based
on that human injection. Similarly, tools like CrewAI allow developers to define hard limits on loop iterations to
prevent agents from spiraling into resource-wasting loops.</t>
            </li>
            <li>
              <t>External Kill-Switch Toolkits: Emerging control architectures, such as the Microsoft Agent Control Toolkit, introduce
decoupled frameworks to monitor and forcefully halt an agent's execution externally when explicit trust, risk, or
compliance boundaries are broken. However, these toolkit-driven intervention signals remain tightly bound to their
native vendor platforms and lack proven interoperability when managing heterogeneous multi-vendor agent topologies
across a production network.</t>
            </li>
          </ul>
        </section>
      </section>
      <section anchor="existing-identity-access-management-approaches">
        <name>Existing Identity &amp; Access Management Approaches</name>
        <t>Modern network agents operate with high degrees of dynamic autonomy, executing complex cross-domain workflows that
alter live network states based on evolving context. Traditional Identity and Access Management (IAM) frameworks
cannot securely map to these systems, as they assume deterministic software processes or static human identities
with pre-provisioned metadata. Existing identity and access mechanisms address this mismatch by introducing
sandboxed runtime environments, short-lived tokens, and behavior-driven trust tracking to defend against malicious
input manipulation and unauthorized tool execution boundaries.</t>
        <t>Implementing Identity &amp; Access Management for Autonomous Agents focuses heavily on sandboxing, adaptive credentialing,
and runtime trust metrics to establish security boundaries and mitigate execution risks early as follows:</t>
        <ul spacing="normal">
          <li>
            <t>Sandboxed Environment Isolation: A prominent current approach to manage agent risk is isolating the agent's
runtime execution loop into secure sandboxes. Operators run the agent in a highly restricted environment so
that it cannot cross structural trust boundaries or compromise broader critical infrastructure, even if the
underlying model is actively manipulated.</t>
          </li>
          <li>
            <t>Static Privilege Control &amp; Short-Lived Tokens: The most widely deployed baseline strategy applies static
least-privilege principles expressed through standard IAM constructs like service accounts and OAuth scopes.
To counter long-term credential leakage risks, best practices utilize short-lived, limited-access Json Web
Tokens (JWTs) with narrow scopes and rapid expirations, restricting tool access to a highly compressed
temporal window.</t>
          </li>
          <li>
            <t>Context-Aware and Dynamic Trust Level Assignment: Advanced emerging mechanisms move past static
configurations by computing dynamic, behavior-driven trust scores in real time. Instead of labeling an
agent as permanently trusted, frameworks (such as the Microsoft Agent Control Toolkit) dynamically
track operational compliance; the score drops instantly upon policy violations, automatically shrinking
permission profiles or restricting the agent's active trust zone.</t>
          </li>
        </ul>
      </section>
      <section anchor="an-audit-architecture-for-cross-domain-interaction-traceability-and-verification">
        <name>An AUDIT Architecture for Cross-Domain Interaction Traceability and Verification</name>
        <t>Modern agents orchestrate complex, long-running workflows that frequently span multiple administrative
boundaries and invoke independent subagents or external tools without active human oversight at each step.
When anomalies or compliance violations occur, local agent logs are insufficient because no single
administrative party possesses full visibility over a multi-domain interaction chain. Traditional
auditing and logging techniques lack the cryptographic primitives required to produce verifiable proof
of execution to external, untrusted entities. The emerging approach in <xref target="I-D.kuehlewind-audit-architecture"/>
addresses these limitations by introducing standard data models and protocol-layer extensions to securely record,
correlate, and verify the multi-domain provenance of agent interactions while enforcing strict user privacy boundaries.</t>
        <t>Implementing <xref target="I-D.kuehlewind-audit-architecture"/> focuses heavily on cross-domain correlation, dynamic authorization
tracking, and independent audit verifiability to track the entire action chain through the following
mechanisms:</t>
        <ul spacing="normal">
          <li>
            <t>Cross-Domain Correlation and Common Identifiers: Rather than relying on isolated execution traces, the
AUDIT framework adapts protocol-layer extensions (such as HTTP headers or context propagation tokens)
to inject a common identifier across administrative boundaries. This allows an independent auditor to
stitch together separate interaction records—spanning user intent, agent-to-agent delegation, and
downstream tool invocations—into a singular, chronologically ordered action chain.</t>
          </li>
          <li>
            <t>Time-Evolving Authorization and Identity Attestation: Unlike traditional static software processes,
an agent's permissions shift non-deterministically based on runtime execution context and delegated
trust boundaries. The framework leverages standardized data representations to capture authorization transitions
as a continuous, time-evolving state, while cleanly differentiating between user, agent instance, and
target service identities across the transaction history.</t>
          </li>
          <li>
            <t>Independent Verifiability and Privacy-Preserving Logging: To transform standard operational telemetry
into a tamper-evident AUDIT record, the architecture composes verifiable building blocks like
attestation (RATS) and transparency logging (SCITT). AUDIT records are designed to be fully verifiable
by a third party that trusts neither the agent nor its operator, utilizing strict selective disclosure
mechanisms to omit sensitive prompt content or personal data while maintaining total cryptographic
integrity.</t>
          </li>
        </ul>
      </section>
    </section>
    <section anchor="gap-analysis-of-the-current-approaches">
      <name>Gap Analysis of the Current Approaches</name>
      <section anchor="gap-in-opentelemetry-for-agent-observability">
        <name>Gap in OpenTelemetry for Agent Observability</name>
        <t>While OpenTelemetry (OTel) is the industry standard for collecting traces, metrics, and logs, it has critical limitations when
applied to Autonomous Agent observability.</t>
        <ul spacing="normal">
          <li>
            <t>Continuous Monitoring and Proactive Evaluation: The fundamental limitation is that OpenTelemetry functions as a passive data
plane for system performance, not a proactive evaluation or guardrail engine for AI behavior. It can track how an application
runs, but it struggles to evaluate what an agent decides based on network topology, network element and network resource
context information, therefore it is hard to track network configuration enforcement confirmation and change impact assessment.</t>
          </li>
          <li>
            <t>Lack Evaluation Metrics: Furthermore, OpenTelemetry only captures the execution process, not objectives, constraints to drive
the agent to perform a task. It lacks native support for bserving metrics such as internal confidence levels associated with
an agent's reasoning logic.</t>
          </li>
          <li>
            <t>Human on the Loop Support: OpenTelemetry originated in cloud-native microservice architectures, its tracing lifecycle
cannot represent asynchronous Human-in-the-Loop (HITL) workflows. Consequently, it provides no mechanism to signal
within a trace that a specific step constitutes a high-risk action, has been suspended, and is currently awaiting
human approval.</t>
          </li>
          <li>
            <t>Accountability Support: Without trajectory record information such as Who executed the network commands, when, based on
what prompts/reasoning, and when, it is difficult to trace responsibility in the event of a network incident, e.g., whether
it is ther responsbility of Autonomous Agent or responsibility of Human operators.</t>
          </li>
        </ul>
      </section>
      <section anchor="gap-in-ai-guardrails">
        <name>Gap in AI Guardrails</name>
        <t>There are many areas where guardrails cannot provide adequate control based on the current capabilities.</t>
        <ul spacing="normal">
          <li>
            <t>Action focus: The majority of the guardrails focus on the text boundary whereas in agentic system the critical
boundary is the action execution, i.e. the point where a tool call, API invocation, or database write reaches a
live system.</t>
          </li>
          <li>
            <t>Multistep execution: Guardrails are typically applied at single-turn boundaries, i.e. they evaluate one input
or one output at a time. Currently, there is no well-defined mechanism for evaluating the control implications
of action sequences, or how a series of individually valid steps may collectively lead to unintended or non-compliant
outcomes. This gap highlights the need for sequence-aware control and intervention mechanisms that can evaluate intent,
track execution context across steps, and assess cumulative impact.</t>
          </li>
          <li>
            <t>Indirect instruction susceptibility: Agents are susceptible to security attacks, particularly those that exploit how
context is constructed and consumed during execution. One prominent class of such attacks is prompt injection, which
takes advantage of a key limitation in current guardrail architectures - i.e. the assumption that malicious or unauthorized
instructions will appear only at the user input boundary. In reality, agentic systems ingest information from multiple
sources, and instructions can be introduced indirectly through retrieved documents (RAG), tool outputs, system messages,
or intermediate reasoning steps. Addressing this limitation requires a shift from boundary-focused guardrails to
context-aware intervention and control mechanisms.</t>
          </li>
          <li>
            <t>Heavy human dependency:  Many guardrail implementations rely on human review for edge cases or escalations, which does not
scale in high-speed or high-volume environments. Also, there is a fine balance required between flexibility of agent execution
and reasoning, with the boundary of execution which is subjective.</t>
          </li>
        </ul>
      </section>
      <section anchor="gap-in-agent-drift-analysis">
        <name>Gap in Agent Drift Analysis</name>
        <t>Agent drift is hard to detect as it seldom produce a failure event, only the effect of the drift can be observed through
continuous monitoring. Current drift management techniques are retrospective and does not intercept the degraded agent
behaviour as it occurs or does not automatically adjust agent policy in response to detected drift, and does not coordinate
drift signals with runtime intervention mechanisms. Agent drift management has similarities to Anomaly management. So a
potential direction is to leverage some of the techniques used in Anomaly management applied to Agents.</t>
      </section>
      <section anchor="gap-in-quality-gates">
        <name>Gap in Quality gates</name>
        <t>Three limitations characterize current implementation of quality gate.</t>
        <ul spacing="normal">
          <li>
            <t>Implementation dependency on framework primitives. Quality gate behaviour is entirely developer-constructed from
framework-specific primitives. There is no standard quality gate interface, no standard evaluation schema, no standard
routing decision vocabulary, and no standard AUDIT record format. A quality gate implemented in LangGraph is architecturally
incompatible with one implemented in CrewAI or AutoGen. They cannot be governed, observed, or audited through common
infrastructure.</t>
          </li>
          <li>
            <t>Absence of external observability. Quality gate evaluations are internal to the agent workflow. No current framework
provides a standardised mechanism for an external observability authority to observe what quality evaluation was performed, what
dimensions were assessed, what score was produced, and why a specific routing decision was made.</t>
          </li>
          <li>
            <t>No central intervention and control. Quality gate outcomes particularly human review and/or rejection decisions
are not connected to a central control infrastructure. A gate that routes to human review pauses execution within the agent
framework, but that pause is not expressed as a standardised intervention signal that a central control authority can monitor,
escalate, or resolve. The gate operates in isolation from the broader management and control stack.</t>
          </li>
        </ul>
      </section>
      <section anchor="gap-in-the-existing-intervention-approaches">
        <name>Gap in the existing Intervention Approaches</name>
        <t>As highlighted above intervention mechanisms exist in primitive and framework-specific forms. They have the following limitations.</t>
        <ul spacing="normal">
          <li>
            <t>Absence of a standardised external interface that allows an authorized authority outside the framework or outside the
agent application to signal intervention and receive a guaranteed response</t>
          </li>
          <li>
            <t>Current practice of intervention (leveraging infrastructure level interventions) is largely binary: either the agent runs or it
does not. It lacks mechanisms that are flexible and applied across spectrum of scenarios - soft redirect, scope restriction,
checkpoint, task suspension, rollback, hard termination</t>
          </li>
          <li>
            <t>When current intervention mechanisms stop an agent , whether through process termination, task cancellation, or API revocation,
they do not systematically preserve the agent's execution state in a form that enables recovery.</t>
          </li>
          <li>
            <t>Current intervention mechanism requires either a human decision or a pre-coded condition to trigger it. There is no mechanism
that continuously monitors agent behaviour against control policies and automatically triggers a proportionate intervention
response when a deviation is detected</t>
          </li>
        </ul>
      </section>
      <section anchor="gap-in-the-existing-identity-and-access-management-approaches">
        <name>Gap in the existing Identity and Access Management Approaches</name>
        <t>From the I&amp;C perspective following are some of the key limitations in incorporating identity and access management in the agents.</t>
        <ul spacing="normal">
          <li>
            <t>Security control mechanisms primarily govern inputs and outputs, but have limited ability to fully interpret or validate the
internal reasoning process of the agent. As a result, reasoning errors or misalignment may go undetected until they take effect
through actions.</t>
          </li>
          <li>
            <t>In federated or multi-agent environments, enforcing consistent security and I&amp;C policies across domains is complex. Differences
in trust models, policies, and enforcement mechanisms can lead to gaps in control.</t>
          </li>
          <li>
            <t>Security standardization for agents is still evolving (e.g. OWASP Top 10 for Agentic Applications provide an emerging taxonomy
of agent-related security risks) but they remain primarily focused on risk identification rather than operational control. At
present, most control mechanisms are tightly coupled to specific frameworks or vendor implementations, leading to fragmented and
non-interoperable approaches.</t>
          </li>
          <li>
            <t>Current frameworks do not provide a consistent approach to handle of delegated trust which is the trust relationship that arises
when an orchestrating agent delegates authority to a sub-agent. For example, when a highly trusted orchestrator assigns a task to
a sub-agent, it is unclear what level of trust the sub-agent should inherit, or what constraints should govern the delegated
authority.</t>
          </li>
        </ul>
      </section>
      <section anchor="gap-in-the-audit-architecture-approach">
        <name>Gap in the AUDIT Architecture Approach</name>
        <t>There are several areas where the AUDIT framework, introduced in <xref target="I-D.kuehlewind-audit-architecture"/>, cannot provide comprehensive
operational control or real-time risk mitigation within autonomous network environments, based on its defined charter and design constraints:</t>
        <ul spacing="normal">
          <li>
            <t>Post-mortem Rather than Pre-Event Interceptive Control: The AUDIT framework is fundamentally optimized for
cryptographic evidence collection and post-event verification across trust domains. It completely lacks protocol-layer
primitives to enforce inline runtime policies or execute real-time task instructions such as risk control,access control
policy precedence control, etc. This means a misbehaving agent generating high-velocity loop mutations cannot be actively
intercepted or suspended before the configuration changes damage the network infrastructure.</t>
          </li>
          <li>
            <t>Absence of In-Event Active State Recovery &amp; Rollback: While AUDIT focuses heavily on recording time-evolving authorization
transitions and chain-of-custody data across distributed workflows, it lacks execution awareness. The framework specifies no
mechanism to checkpoint active agent states, preserve runtime memory contexts, or trigger granular rollbacks (such as reverting
a partial network configuration to a last known safe state) interrupt execution, human escalation,failure correction when a
cross-domain policy violation occurs.</t>
          </li>
          <li>
            <t>Exclusion of Internal Reasoning Evaluation: The primary limitation of the AUDIT approach is its strict focus on external, observable
behaviors and boundary interaction states. It explicitly excludes the evaluation of underlying Large Language Models (LLMs) and internal
reasoning process. Consequently, structural reasoning loops related to planning, self-optimization, chain of thoughts,memory retrieval
remain completely invisible within the logged trajectory record.</t>
          </li>
        </ul>
      </section>
    </section>
    <section anchor="standardization-area">
      <name>Standardization Area</name>
      <t>This section outlines key areas where standardization is required to support the design, implementation, and operation of Network
Management Agent Observability, Intervention and Control in Agent Fabric networks. In the Agent Fabric Network,
- Two or multiple scenario specific Autonomous Agents such as network management agents can work together to support multi-scenario autonomy or close loop management.
- Two or muitiple scenario specific Autonomous Agents such as network management agents can work together to support cross domain collaboration.
- Two or mutiple sceanrio specific Autonomous Agents such as network management agents can work together to support collaboration between service layer and network layer.
the agent gateway can be used to collect metric, log, audit information from each Autonomous Agent.</t>
      <t>The intent is to identify foundational areas that require align with network management technologies developed in IETF OPS Area and drive network
automation moving toward AI Driven Network Operation.</t>
      <ul spacing="normal">
        <li>
          <t>Agent Observability, Intervention and Control Network Management Architecture: Developing or selecting a framework for enforcing boundaries,
detecting, evaluating, interrupting, correcting, and recovering from agent behavior within operational and compliance boundaries.</t>
        </li>
        <li>
          <t>Agent Observability Data Model: Enabling network behavioral assessment through analysis of observed operational network data
(logs, metrics, traces, etc.)</t>
        </li>
        <li>
          <t>Human and Agent Interaction protocol for Human Escalation/Intervention, Agent Intervention and Control</t>
        </li>
      </ul>
    </section>
    <section anchor="security-considerations">
      <name>Security Considerations</name>
      <t>The security considerations applicable to Network Digital Twin and Agentic AI based Architecture for AI driven Network Operations
<xref target="I-D.wmz-nmrg-agent-ndt-arch"/> are also applicable to this document.</t>
    </section>
    <section anchor="iana-considerations">
      <name>IANA Considerations</name>
      <t>This document has no IANA actions.</t>
    </section>
  </middle>
  <back>
    <references anchor="sec-combined-references">
      <name>References</name>
      <references anchor="sec-normative-references">
        <name>Normative References</name>
        <reference anchor="RFC2119">
          <front>
            <title>Key words for use in RFCs to Indicate Requirement Levels</title>
            <author fullname="S. Bradner" initials="S." surname="Bradner"/>
            <date month="March" year="1997"/>
            <abstract>
              <t>In many standards track documents several words are used to signify the requirements in the specification. These words are often capitalized. This document defines these words as they should be interpreted in IETF documents. This document specifies an Internet Best Current Practices for the Internet Community, and requests discussion and suggestions for improvements.</t>
            </abstract>
          </front>
          <seriesInfo name="BCP" value="14"/>
          <seriesInfo name="RFC" value="2119"/>
          <seriesInfo name="DOI" value="10.17487/RFC2119"/>
        </reference>
        <reference anchor="RFC8174">
          <front>
            <title>Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words</title>
            <author fullname="B. Leiba" initials="B." surname="Leiba"/>
            <date month="May" year="2017"/>
            <abstract>
              <t>RFC 2119 specifies common key words that may be used in protocol specifications. This document aims to reduce the ambiguity by clarifying that only UPPERCASE usage of the key words have the defined special meanings.</t>
            </abstract>
          </front>
          <seriesInfo name="BCP" value="14"/>
          <seriesInfo name="RFC" value="8174"/>
          <seriesInfo name="DOI" value="10.17487/RFC8174"/>
        </reference>
      </references>
      <references anchor="sec-informative-references">
        <name>Informative References</name>
        <reference anchor="MCP" target="https://modelcontextprotocol.io/">
          <front>
            <title>Model Context Protocol</title>
            <author>
              <organization/>
            </author>
            <date year="2024" month="November"/>
          </front>
        </reference>
        <reference anchor="A2A" target="https://google-a2a.github.io/A2A/#/documentation?id=agent2agent-protocol-a2a">
          <front>
            <title>Agent2Agent (A2A) protocol</title>
            <author>
              <organization/>
            </author>
            <date year="2025" month="April"/>
          </front>
        </reference>
        <reference anchor="IG1507" target="https://projects.tmforum.org/wiki/pages/viewpage.action?pageId=411641744">
          <front>
            <title>TM Forum, IG1507 Intervention and Control for Agentic Operation V1.0.0 DRAFT</title>
            <author>
              <organization/>
            </author>
            <date year="2026" month="May"/>
          </front>
        </reference>
        <reference anchor="IG1251G" target="https://projects.tmforum.org/wiki/pages/viewpage.action?pageId=401824956">
          <front>
            <title>TM Forum, IG1251G IP Network AN Level 4 Agentic Architecture for Multi-Scenario Autonomy</title>
            <author>
              <organization/>
            </author>
            <date year="2026" month="May"/>
          </front>
        </reference>
        <reference anchor="OpenTelemetry" target="https://opentelemetry.io/">
          <front>
            <title>OpenTelemetry (OTel)</title>
            <author>
              <organization/>
            </author>
            <date>n.d.</date>
          </front>
        </reference>
        <reference anchor="I-D.kuehlewind-audit-architecture">
          <front>
            <title>An Architecture for Auditing Agent Delegation and Interactions</title>
            <author fullname="Mirja Kühlewind" initials="M." surname="Kühlewind">
              <organization>Ericsson</organization>
            </author>
            <author fullname="Henk Birkholz" initials="H." surname="Birkholz">
              <organization>Fraunhofer SIT</organization>
            </author>
            <date day="7" month="September" year="2026"/>
            <abstract>
              <t>   This document describes an architecture for auditing of agent-driven
   interactions on the Internet.  Autonomous and semi-autonomous
   software agents, including those based on artificial intelligence,
   increasingly act on behalf of users, organizations, and services.
   Existing auditing mechanisms often capture isolated system events but
   do not consistently represent delegation relationships, user intent,
   or evolving authorization.  In agent-driven systems, auditability
   requires linking intent, delegation, authorization, and execution.
   The proposed architecture enables this through distributed audit
   record generation, propagation of audit context, optional
   attestation, and additional logging for transparency.

              </t>
            </abstract>
          </front>
          <seriesInfo name="Internet-Draft" value="draft-kuehlewind-audit-architecture-01"/>
        </reference>
        <reference anchor="I-D.wmz-nmrg-agent-ndt-arch">
          <front>
            <title>Network Digital Twin and Agentic AI based Architecture for AI-driven Network Operations</title>
            <author fullname="Qin Wu" initials="Q." surname="Wu">
              <organization>Huawei</organization>
            </author>
            <author fullname="Cheng Zhou" initials="C." surname="Zhou">
              <organization>China Mobile</organization>
            </author>
            <author fullname="Luis M. Contreras" initials="L. M." surname="Contreras">
              <organization>Telefonica</organization>
            </author>
            <author fullname="Sai Han" initials="S." surname="Han">
              <organization>China Unicom</organization>
            </author>
            <author fullname="Yong-Geun Hong" initials="Y." surname="Hong">
              <organization>Daejeon University</organization>
            </author>
            <date day="28" month="August" year="2026"/>
            <abstract>
              <t>   A Network Digital Twin (NDT) provides a network emulation tool usable
   for different purposes such as scenario planning, impact analysis,
   and change management.  Agentic AI enables dynamic goal-driven
   execution and adaptive behavior and closed-loop autonomy.  By
   integrating a NDT into network management together with the Agentic
   AI, it allows the network management activities to take user intent
   or service requirements as input, automatically assess, model, and
   refine optimization strategies under realistic conditions but in a
   risk-free environment.  Such environment that operates to meet these
   types of requirements is said to have AI-driven network operations.

   AI-driven network operations brings together existing technologies
   such as Agentic AI and NDT which may be seen as the use of a toolbox
   of existing components enhanced with a few new elements.

   This document describes an architecture for AI-driven network
   operations and shows how these components work together with NDT and
   Agentic AI capabilities.

              </t>
            </abstract>
          </front>
          <seriesInfo name="Internet-Draft" value="draft-wmz-nmrg-agent-ndt-arch-06"/>
        </reference>
      </references>
    </references>
    <?line 791?>

<section numbered="false" anchor="acknowledgments">
      <name>Acknowledgments</name>
      <t>The authors of this document would also like to thank Benoit Claise, Daniele Ceccarelli for review and comments.</t>
    </section>
    <section anchor="contributors" numbered="false" toc="include" removeInRFC="false">
      <name>Contributors</name>
      <contact fullname="Swamynathan Balasundaram">
        <organization>Nokia</organization>
        <address>
          <email>swamynathan.b@nokia.com</email>
        </address>
      </contact>
      <contact fullname="Qiufang Ma">
        <organization>Huawei</organization>
        <address>
          <email>maqiufang1@huawei.com</email>
        </address>
      </contact>
      <contact fullname="Giuseppe Fioccola">
        <organization>Huawei</organization>
        <address>
          <email>giuseppe.fioccola@huawei.com</email>
        </address>
      </contact>
    </section>
  </back>
  <!-- ##markdown-source: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-->

</rfc>
