| Internet-Draft | DAWN Use Cases | November 2026 |
| King, et al. | Expires 20 May 2027 | [Page] |
This document describes use cases and applicability for Discovery of Agents With Names (DAWN). It illustrates how clients discover AI resources and obtain the minimum information needed for subsequent interaction within a local network, within an organisation, or between cooperating organisations with trust relationships.¶
This document does not define a discovery protocol, a registration procedure, a selection algorithm, or an agent-to-agent communication protocol.¶
This note is to be removed before publishing as an RFC.¶
Status information for this document may be found at https://datatracker.ietf.org/doc/draft-kay-dawn-use-cases/.¶
Discussion of this document takes place on the Discovery of Agents With Names Working Group mailing list (mailto:dawn@ietf.org), which is archived at https://mailarchive.ietf.org/arch/browse/dawn/. Subscribe at https://www.ietf.org/mailman/listinfo/dawn/.¶
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.¶
Copyright (c) 2026 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
Artificial Intelligence (AI) systems may depend on resources whose capabilities, locations, or communication methods are not configured in advance. Discovery allows a client to locate a resource and obtain the minimum information needed to attempt a subsequent interaction.¶
Discovery of Agents With Names (DAWN) initially focuses on AI agents and their capabilities, including discovery of other resources used in AI systems. This document uses "entity" for a discoverable component in that context. Such entities may include agents, tools, services, tasks, data sources, models, inference services, workloads, brokers, and application endpoints. Skills may be discoverable attributes of those entities. Reuse in other discovery contexts is desirable, but does not extend the initial scope to general resource discovery.¶
The DAWN problem statement [I-D.akhavain-moussa-dawn-problem-statement] provides background for this work. These use cases are intended as a common foundation for the DAWN requirements [I-D.king-dawn-requirements] and for the Discovery Architecture, which will define the minimum discovery information. The mapping between requirements and use cases is maintained in the requirements document, not here.¶
This document first presents four scenarios in Section 5. It then draws out the discovery characteristics and common pattern they share, and defines the following categories of entity discovery:¶
Capability-Oriented Discovery, where an entity needs to discover another entity that can provide a function or capability.¶
Resource-Oriented Discovery, where discovery identifies resources used in AI systems.¶
Administrative Scope Extensions, where discovery crosses organisational, delegation, or tenancy boundaries within the applicable trust relationships.¶
Operational Discovery, where discovery supports operation, audit, troubleshooting, compliance, or automation.¶
The first two categories are base cases. The others describe common extensions or deployments of those base cases.¶
Other Internet-Drafts describe agentic AI discovery use cases and communication protocol requirements, including [I-D.mozley-aidiscovery], [I-D.agentic-ai-usecases-requirements], and [I-D.scrm-aiproto-usecases].¶
This document describes discovery within a local network, within an organisation, or between cooperating organisations with trust relationships over the public Internet. It does not describe the full lifecycle of an entity or establish those trust relationships.¶
The following are in scope:¶
discovery of a specific named AI resource;¶
discovery of candidate AI resources by capability or other discovery properties within the applicable deployment;¶
discovery of tasks as resources in an AI system;¶
discovery initiated by an agent, workload, service, application, or human operator;¶
discovery of supporting resources used by AI agents and applications;¶
exchange of the minimum discovery information, directly or by indirection, before selection, invocation, or communication can take place.¶
The following are out of scope:¶
registration of entities for discovery;¶
authentication or attestation of registrations;¶
design, definition, or governance of naming systems;¶
identity management of AI agents, tools, and skills;¶
trust management and trust evaluation beyond protecting content exchanged between discovery participants;¶
selection among candidate entities returned by discovery;¶
exchange of capability information beyond the minimum discovery information defined by the Architecture work;¶
capability negotiation after discovery;¶
all application communication beyond the discovery process;¶
task orchestration;¶
standalone certificate or key lookup services; and¶
AI resource indexing across the wider Internet, general search, ranking, or marketplace discovery.¶
Clients may query an organisation's catalogue or a cooperating discovery service within these boundaries. They may follow references to obtain minimum discovery information. Fetching a richer capability description is a subsequent application exchange, even when discovery supplies its reference.¶
Terminology for DAWN is defined in [I-D.farrel-dawn-terminology].¶
Attention is drawn to the following specific terms defined in [I-D.farrel-dawn-terminology] that are key to this document:¶
discovering entity¶
discovered entity¶
discoverable object¶
discovery information¶
minimum discoverable information¶
The client initiating discovery is the discovering entity. In this document, "minimum discovery information" refers to the minimum discoverable information whose content is to be defined by the Architecture requirements.¶
The scenarios in Section 5 also use the following implementation term:¶
Software that supports an AI agent's execution, such as interaction with models, invocation of tools, management of context and state, and application of execution controls. It may support discovery on behalf of the agent. This document does not specify its internal operation. Open-source examples of an agent harness include OpenClaw and Hermes.¶
The categories in Section 8 apply in three settings:¶
| Deployment | Illustrative discovery need | Applicability condition |
|---|---|---|
| Local network | An assistant finds a nearby inference service advertised on its local network. | Discovery visibility is local; using the service remains subject to application permissions. |
| Within an organisation | An enterprise assistant finds an internal data source or tool. | Publication and access follow the organisation's existing administrative and trust arrangements. |
| Between cooperating organisations | An assistant finds a partner's calendar-coordination agent. | The organisations have a trust relationship; discovery finds resources within that relationship. |
Organisations publish independently and may use delegation or federation under existing trust arrangements. Knowing an organisation's name does not establish trust. Conversely, a partner's agents and endpoints may be unknown even when the organisational relationship is established.¶
DNS and mDNS are possible discovery mechanisms; others may also be suitable. Protocol work should reuse existing IETF mechanisms where possible.¶
The four scenarios in this section ground the rest of the document. The first follows a personal assistant coordinating an appointment across an organisational boundary. The second follows a media orchestration agent provisioning edge compute within a broadcaster and its cooperating providers. The third follows the same personal assistant discovering the agents of an unfamiliar building from the local network it has joined. The fourth illustrates infrastructure-assisted discovery in a managed enterprise network. The discovery characteristics, pattern, and categories in the following sections draw on them.¶
Alice asks her AI assistant, Hermes, to review her schedule and arrange lunch with Bob at another organisation, including a restaurant table for two near the office. Later, she asks to move both the appointment and the reservation to a different day.¶
The example combines capability and resource discovery across an organisational boundary, then shows how Hermes can reuse what it has learned.¶
Hermes Agent provides the agent harness for Alice's assistant. The running assistant, called Hermes here, is the discovering entity. Alice talks to it through a conversational interface. Her calendar service is already configured; Hermes can read it but cannot change it.¶
Alice's and Bob's organisations are assumed to cooperate under an established trust relationship. Their calendar agents have not previously interacted. Hermes knows Bob's organisation but not its calendar-agent endpoints. Bob's agent can coordinate appointments and create events, subject to permission.¶
The restaurant-booking service is assumed to be available through a cooperating provider with an established trust relationship. OpenTable is used as an illustrative provider; the example assumes a suitable agent-accessible interface and does not claim that a particular integration or trust arrangement exists.¶
Alice is assumed to supply enough context to distinguish the intended Bob.¶
Alice asks Hermes to find a time for lunch with Bob on Thursday and book a table for two near the office. Hermes reads her calendar, then needs to find a calendar-coordination agent at Bob's organisation. It knows the organisation and required capability, and may have constraints on the communication protocol.¶
Hermes locates the organisation's discovery entry point and queries its catalogue for calendar-coordination agents. The result identifies candidates and their contact information. DNS could provide the entry point; the example does not depend on a particular bootstrap or catalogue protocol.¶
A candidate may serve the whole organisation or particular users. Hermes needs to determine whether it can handle the request concerning Bob. The application then selects a candidate, obtains any necessary authorisation, and exchanges proposed times. Bob's agent creates the event in his calendar. Hermes still has no permission to write to Alice's calendar; invitation handling is an application matter.¶
Hermes obtains free/busy information through the calendar services. Alice's calendar was already known, so reading it required no discovery. It discovers any additional data source it needs before querying that source.¶
With a suitable time identified, Hermes proceeds as follows:¶
Discover the restaurant-booking service through the cooperating provider's discovery entry point. The result identifies the booking capability and how to contact it.¶
Query the service for restaurants near the office with a table for two at the proposed time. Availability, price, and booking conditions come from the service's application interface.¶
Present a suitable option to Alice and ask her to confirm the restaurant, time, and any booking conditions.¶
On confirmation, request the reservation through the service and retain its booking reference. Report whether the reservation succeeded.¶
The discovery target is the booking service. Restaurant search, table availability, and reservation are subsequent application activities. The table is reserved through the service, not through DAWN.¶
The following exchanges show the restaurant-booking part of the example. The discovery steps also apply to finding Bob's calendar agent, using his organisation's discovery entry point instead of the booking provider's.¶
Alice Hermes Provider's Booking
discovery service
| | | |
|--Lunch request--> | |
| | | |
Discovery
| |-Booking capability--> |
| <-Name, URI, protocol-| |
| | | |
Application exchange
| |----------Time, area, party size----------->
| <--------Options, price, conditions---------|
<-Booking option--| | |
|-----Confirm-----> | |
| |----------Reserve selected table----------->
| <---------Result, booking reference---------|
<-Booking result--| | |
| | | |
¶
Alice asks Hermes to move the lunch to Friday. Hermes has cached discovery information for both the calendar agent and the booking service, including their names, endpoints, protocols, and available freshness information. If these remain usable, it can reconnect without repeating the catalogue queries.¶
An expired record or changed endpoint may require a refresh. If the agent cannot be found or no longer offers calendar coordination, Hermes can return to the organisation's discovery entry point. A temporary connection failure alone does not establish that the agent has been withdrawn.¶
The application may also need to retrieve a changed capability card and reassess the agent. Finding a replacement does not establish that it holds the original booking or has permission to change it.¶
Hermes checks Friday's calendar availability and asks the booking service whether the reservation can be moved. It presents the revised arrangements and any changed conditions to Alice for confirmation, then requests both changes through the respective services. If only one succeeds, Hermes reports the outstanding change to Alice. Discovery does not make the two updates a single transaction.¶
The services below are contacted separately. If cached information is still usable, the discovery exchanges are skipped. Availability checks and booking changes use application interfaces.¶
Alice Hermes Discovery Calendar /
entry points booking
services
| | | |
|-Move to Friday--> | |
| | | |
Reuse cached discovery information if still valid.
If refresh or rediscovery is needed:
| |-Name or capability--> |
| <---Contact details---| |
| | | |
Application exchanges with each service
| |--Friday availability / booking reference-->
| <------Options and changed conditions-------|
<--Revised plan---| | |
|-----Confirm-----> | |
| |-----Change appointment / reservation------>
| <-----Separate results for each change------|
<-----Outcome-----| | |
| | | |
¶
The first requests identify an organisation and a capability: calendar coordination or restaurant booking. Reconnection starts with a known agent or service. In either case, useful results include:¶
the entity's name, type, and a service or capability identifier;¶
reachability information and supported communication methods; and¶
references used to obtain that minimum information through indirection.¶
The Architecture defines the minimum field set. A reference may lead to further discovery information or to richer application information. Fetching a capability card beyond that minimum, detailed policy, live calendar data, or table availability is outside discovery. Booking references are application state, separate from cached discovery information.¶
All cross-organisation discovery information is integrity-protected under the assumed trust arrangements, including information obtained through references. Hermes checks its origin and freshness. An authentic publication does not prove that an agent may act for Bob. A version comparison can reveal change, but does not establish authenticity.¶
The discovery query may reveal Alice's interest in Bob's organisation and its scheduling service. Calendar contents and personal details can be restricted to the subsequent authorised exchange. Discovery records may also need access controls.¶
For troubleshooting, operators may need to know which source Hermes used, whether it used cached information, and why it repeated discovery. Logs can record these decisions, subject to privacy and retention controls.¶
A broadcaster offers a personalised live programme [object-based media delivery]. Viewers may choose different audio, graphics, or accessibility features. Some devices cannot combine these media objects themselves, so nearby edge compute renders part or all of the programme for them.¶
Before the broadcast, an operator asks a media orchestration agent to arrange enough compute for the expected audience, balancing delay, cost, and energy use. The agent needs to find the services and information sources that can help it do this.¶
The media orchestration agent coordinates the request. A compute allocation agent arranges resources at suitable sites, using an optimisation tool to compare options. Data services provide audience forecasts and information about compute and network conditions.¶
These entities operate within the broadcaster's organisation or cooperating providers with established trust relationships. The discovery entry points are known, but individual service endpoints may not be.¶
Agentic Resource Discovery (ARD) is a possible interface to a provider's catalogue in this scenario. It provides a common way to describe and search agents, tools, and services. A provider could expose its compute allocation and telemetry services through ARD, allowing the media agent to find candidates without preconfiguring every service endpoint or loading every service description into its context. Here, queries would be limited to the organisation or cooperating providers. ARD is an illustrative option; protocol choice remains open. Resource allocation and invocation continue through the service's own interfaces.¶
The media agent queries the relevant discovery entry points for a compute allocation capability and any required data services. Results identify the entities and how to contact them.¶
The allocation agent discovers suitable site controllers and media-processing services, including GPU rendering where needed. It obtains their names, service identifiers, endpoints, and supported communication methods.¶
Through the services' application interfaces, the agents obtain audience forecasts, current capacity, and cost, energy, and delay estimates. The optimisation tool uses these inputs and the operator's priorities to propose a placement.¶
Subject to the operator's policy and authorisation, the allocation agent requests provisioning through the site controllers. It reports the result, or explains why the available resources cannot meet the request.¶
The diagram summarises discovery and the later application exchanges. The agents contact each service separately.¶
Media / allocation Discovery Data services / agents entry points site controllers | | | |--Required capability---->| | |<--Name, URI, protocol----| | | | | |--Forecast / resource information request-------->| |<--Demand, capacity, cost, energy, delay----------| | | | Compare options using the optimisation tool | | | |--Provision selected resources------------------->| |<--Result / resource references-------------------|¶
Discovery provides the information needed to contact an entity. Comparing sites, reserving compute, steering traffic, and delivering media are application activities outside DAWN.¶
Audience demand or site availability may change. The agents can reuse valid discovery information and query current conditions through the data services. Changed or expired discovery records may require a refresh or a new query. Moving workloads or adding capacity remains an application decision.¶
Discovery information is integrity-protected, including across provider boundaries, and its origin and freshness are checked. Finding a controller does not grant permission to provision resources. Viewer details and detailed capacity data are restricted to authorised application exchanges.¶
Alice visits Bob's office for a meeting. The building is shared by several organisations and run by a facilities operator. On arrival, she joins the building's guest Wi-Fi network and asks Hermes to check her in and show her the way to the meeting room.¶
The example shows discovery that starts from the network a client has joined, rather than from a name. The same applies to visitors at a conference venue, hotel, hospital, airport, or stadium.¶
Hermes, Alice's assistant from Section 5.1, runs on her phone and is the discovering entity. The facilities operator runs a reception agent, which checks visitors in and notifies their hosts, and a wayfinding agent, which gives directions inside the building. Both are advertised on the guest network.¶
Hermes knows neither the operator's name nor any catalogue it publishes, and has no prior relationship with it. The operator's agents are not published at the discovery entry point of Bob's organisation. The network Alice has joined is the only starting point Hermes has.¶
Joining the network is outside discovery. Using the discovered agents remains subject to application permissions.¶
With Alice's phone attached to the guest network, Hermes proceeds as follows:¶
Request the agents advertised on the network it has joined. The request names no organisation, entity, or capability: the network bounds it, as the organisation's discovery entry point bounds the requests in Section 5.1. DNS-based Service Discovery over multicast DNS could provide this; the example does not depend on a particular discovery protocol.¶
Receive, for each advertised agent, its name, its communication protocol, and how to contact it, possibly with a reference to a richer description. The names, such as "Reception" and "Wayfinding", are meant to be read by people. An agent may support more than one protocol, for example A2A and MCP, in which case Hermes uses one that it supports.¶
Present the names to Alice, who picks the agents to use. Alternatively, Hermes can retrieve the referenced descriptions and decide which agents fit her request. Either way, the choice is made after discovery.¶
Check Alice in through the reception agent, which notifies Bob, then ask the wayfinding agent for directions to the meeting room.¶
The discovery target is the set of agents the building advertises. Choosing among them, checking in, and finding the way are subsequent application activities. To notify Bob, the reception agent may itself discover an agent at the discovery entry point of his organisation, as in Section 5.1.¶
Alice Hermes Building's Reception /
guest network wayfinding
agents
| | | |
|---Check me in---> | |
| | | |
Discovery within the joined network
| |--Advertised agents--> |
| <-Name, URI, protocol-| |
| | | |
<---Agent names---| | |
|---Pick agents---> | |
| | | |
Application exchanges with each agent
| |--------------Check in Alice--------------->
| <---------Checked in, host notified---------|
| |-----------Route to meeting room----------->
| <----------------Directions-----------------|
<---Directions----| | |
| | | |
¶
While Alice is in the building, Hermes can reuse the discovery information it has obtained. Agents may appear or be withdrawn during her visit; an agent that serves a single event, for example, is withdrawn when the event ends. A changed or withdrawn agent may require a refresh or a new request.¶
The information applies only to the network on which it was obtained. When Alice leaves the building, Hermes stops relying on it; at another building it discovers again. If Alice returns the next day, Hermes cannot assume that the same agents are still advertised.¶
In a large building or campus, the guest network may span several network segments. Alice expects the same agents to be discoverable wherever she is attached, so the operator may make the same discovery information available across those segments. How it does so is a deployment matter.¶
The operator also decides which agents are visible on which network. Agents for staff, such as building-control agents, are not advertised on the guest network.¶
The request identifies no organisation, entity, or capability; the joined network bounds it. As in Section 5.1, useful results include the entity's name, reachability information and supported communication methods, and references used to obtain further information. The starting point differs, but discovery on a local network adds no fields to the minimum discovery information.¶
Names are most useful when they are meaningful to people, because Alice may choose among the agents directly. Evaluating what an agent can do, beyond the minimum discovery information, is outside discovery.¶
Like a well-run enterprise network, the building's network can offer some protection against rogue advertisements. The operator can isolate guest devices from one another, or accept advertisements on the guest network only from its own infrastructure, so that another visitor cannot advertise an agent named "Reception". These measures complement integrity protection and an origin check of the discovery information, as in the other scenarios, but do not replace them: not every network is run this way, and locality alone does not establish trust.¶
An authentic advertisement does not entitle an agent to Alice's personal details. Checking in shares her name and Bob's with the operator, in an application exchange that Alice authorises.¶
A request bounded by the joined network names no organisation or capability, so it reveals little about Alice's intent, although other devices on the network may observe it. Advertisements on the guest network are visible to every visitor, so the operator publishes there only what all visitors may see, without exposing its internal structure or its tenants.¶
An enterprise operates multiple campuses and edge-compute sites. Field engineers use an AI Operations Assistant to diagnose equipment and network problems. Troubleshooting Agents may be deployed at site, regional-edge, or shared-cloud locations.¶
The Operations Assistant may need to obtain the information required to connect to an identified Troubleshooting Agent or troubleshooting service. In a managed deployment, some information useful to discovery may already be available from enterprise infrastructure. For example, hosting infrastructure may know the resource's current endpoint and communication protocol, while network infrastructure may know the requester's current site, tenant, VPN, edge region, or other deployment context.¶
This infrastructure-provided information may contribute to the minimum discovery information used by the discovery mechanism. How the information is obtained, maintained, or supplied is deployment-specific and outside the scope of this scenario.¶
The Operations Assistant is the discovering entity used by the field engineer. A Troubleshooting Agent or troubleshooting service is the discoverable entity.¶
The enterprise hosting environment maintains current discovery-relevant information about the resource, such as its identifier, externally reachable endpoint, and supported communication protocol.¶
The enterprise network also has information about the Operations Assistant's current administrative or network context, such as site, tenant, VPN, network slice, or edge region. The Operations Assistant does not need to understand or explicitly signal these infrastructure-specific details. Where useful, an abstract representation of this context may form part of the minimum discovery information.¶
These forms of infrastructure assistance correspond to the Resource State Hook and Attachment Context Hook described in [I-D.dunbar-dawn-infra-assisted-discovery]. Resource registration, deployment-state maintenance, network-context derivation, and enterprise-policy configuration are outside the scope of this scenario.¶
An engineer at Site A asks the Operations Assistant to diagnose a problem.¶
An application or enterprise function identifies the Troubleshooting Agent or service to be contacted. DAWN discovery is then used to obtain the minimum information needed to connect to that resource.¶
For example, the discovery input may identify:¶
Troubleshooting-Agent.Site-A¶
At the same time, infrastructure may have requester context such as:¶
Site A Engineering VPN Operations tenant¶
and current resource state such as:¶
Endpoint: https://troubleshoot-a.example Protocol: A2A¶
Requester context may help determine the appropriate discovery domain, resolver view, local discovery scope, or bootstrap entry point. Current resource state may help ensure that the returned endpoint and protocol information are current.¶
Conceptually:¶
Resource State Hook
|
v
Requester -----> DAWN Discovery -----> Discovery Information
^
|
Attachment Context Hook
¶
The requester context is used to assist bootstrap discovery; it is not used by DAWN to compare or select among Troubleshooting Agents. Resource identification, capability evaluation, selection, authorisation, diagnosis, telemetry access, and remediation occur outside the DAWN discovery process.¶
The same model applies to remote troubleshooting. An engineer attached at Site A may be troubleshooting equipment at Site B. An application or enterprise function identifies the relevant Site-B resource, while DAWN obtains the information needed to connect to it. The requester's Site-A context may still help determine the appropriate discovery domain or resolver view.¶
Requester context and resource state may change independently.¶
If the engineer moves from Site A to Site B, a subsequent discovery may use a different discovery domain, resolver view, local discovery scope, or bootstrap entry point.¶
Similarly, a Troubleshooting Agent may move to another hosting environment or receive a new externally reachable endpoint while retaining the same logical identity.¶
Infrastructure-assisted discovery allows subsequent discovery to use current requester context and resource state without requiring the Operations Assistant to understand how the underlying deployment changed.¶
How mobility, reassociation, endpoint changes, or policy mappings are detected and maintained is outside DAWN.¶
Infrastructure-provided information used for discovery needs appropriate integrity, origin, and freshness protection.¶
Requester context may be derived from enterprise-sensitive information such as tenant identifiers, VPNs, network slices, edge locations, or internal topology. An abstract requester context can allow these details to remain hidden from the discovering entity and avoid encoding them directly in the discovery request.¶
Infrastructure assistance does not authorise use of the discovered AI resource or access to enterprise systems. Authentication, authorisation, trust evaluation, and subsequent interaction occur after discovery.¶
The scenarios in Section 5 share a common question: what must a client discover before it can attempt interaction with an entity? The following characteristics are drawn from them.¶
Discovery may be initiated by autonomous software or by a human operator. Alice's assistant discovers on her behalf in Section 5.1; an operator's request leads agents to discover further services in Section 5.2.¶
The discovered target may be a specific named entity or a set of entities determined by properties or classification. The assistant first looks for a calendar-coordination capability at a known organisation, and later reconnects to the specific agent it found. In Section 5.4, the target is an already identified Troubleshooting Agent.¶
The starting point of discovery varies: an organisation's name in Section 5.1, known discovery entry points in Section 5.2, and only the network the client has joined in Section 5.3. The minimum discovery information returned is the same in each case. In Section 5.4, enterprise infrastructure helps determine the starting point from the requester's attachment context.¶
Capability is often an important discovery input, but it is not the only one. Organisation, jurisdiction, locality, protocol, policy, and responsible party also appear in several use cases.¶
The minimum discovery information may change over time. Freshness and caching therefore matter even when detailed runtime properties are obtained later through application interaction. The scenarios reuse cached discovery information and refresh it when it has expired or changed, while calendar contents, table availability, and site capacity are obtained from the services themselves.¶
Discovery may need to account for entities that are dynamic, mobile, or available only in a particular context. Agents are advertised and withdrawn during a visit in Section 5.3; the engineer and the Troubleshooting Agent may each move in Section 5.4.¶
Trust in discovery information is different from trust in the discovered entity. A discovery mechanism can help protect discovery metadata, but it does not decide whether the discovered entity should be used. An authentic record for a calendar agent does not show that the agent may act for Bob, and finding a site controller does not grant permission to provision resources.¶
Brokers and aggregators may serve an organisation or cooperating group. Discovery supports independent publication rather than requiring a single central registry or an index of resources across the wider Internet.¶
The scenarios in Section 5 follow a common pattern.¶
A discovering entity has a task, intent, policy requirement, or operational need.¶
The discovering entity needs to find an entity, or class of entities, that can satisfy that need.¶
The discovering entity forms a discovery request using information such as a name, organisation, entity type, capability, location, jurisdiction, communication protocol, or policy constraint.¶
The discovery mechanism returns information about one or more candidate entities.¶
The discovering entity uses that information to decide whether selection, authorisation, capability exchange, or communication should be attempted.¶
The last step is outside the scope of discovery. However, discovery has to return enough information for this step to be possible.¶
In Section 5.1, Alice's lunch request creates the need (step 1) for a calendar-coordination agent at Bob's organisation (step 2). The assistant queries that organisation's discovery entry point by capability (step 3) and receives candidates with their contact information (step 4). Choosing a candidate, obtaining authorisation, and exchanging proposed times are step 5. In Section 5.2, the same steps are repeated for the compute allocation capability, the data services, and the site controllers. In Section 5.3, the request in step 3 names no organisation, entity, or capability: the network the assistant has joined bounds it. In Section 5.4, the request in step 3 names an identified resource, and infrastructure-provided requester context helps determine the discovery domain or entry point used.¶
The Architecture determines which query properties belong in minimum discovery information. Other constraints are evaluated by the application.¶
Discovery may take several steps. A client may follow a reference, refresh a cached result, or repeat a query when the information is no longer sufficient.¶
The categories generalise the scenarios in Section 5. Each category includes assumptions, discovery impacts, and examples within the settings in Section 4.¶
A client needs to find an entity that provides a particular function or capability, such as an agent, tool, task, service, or model-serving endpoint.¶
The client can express the required function using the discovery properties supported by the architecture. A service type or capability identifier may be sufficient. The result identifies the function and how to reach the entity, and may include a reference to a capability card.¶
Capability-oriented discovery suggests support for discovery by function, capability, skill, entity type, and protocol. The returned discovery information needs to include enough information to allow later selection, authorisation, capability exchange, or communication to be attempted.¶
A scheduling assistant needs to find a calendar-coordination agent at a partner organisation and learn how to contact it. This is the scenario in Section 5.1.¶
An agentic workflow decomposes a user request into several subtasks. The workflow needs to discover agents, tools, services, or tasks that can participate in the work. Discovery provides candidate entities and their published properties; the workflow then performs selection and orchestration outside DAWN.¶
A client needs to discover resources supporting an AI system, such as data sources, knowledge bases, models, inference services, or compute services.¶
This discovery category assumes that the discovered resource may have properties that are partly static and partly dynamic. For example, a dataset description may be relatively stable, while freshness or access policy may change. A compute resource may have stable hardware properties but rapidly changing availability, load, price, or locality.¶
It also assumes that some discovery information may be sensitive. Data classification, model provenance, jurisdiction, permitted use, and operational capacity may need visibility controls.¶
Discovery returns the minimum information needed to locate and contact the resource. It may return a pointer to an application service for current state, detailed policy, or authorisation. Accessing an already known data source to read its contents is an application query, not resource discovery.¶
Properties such as jurisdiction, format, freshness, and access method may motivate discovery requirements. The Architecture work determines which belong in the minimum discovery information. Richer descriptions, live availability, and detailed policy information are obtained outside the discovery process.¶
An enterprise agent needs to find a data source for retrieval-augmented generation. Discovery identifies a relevant internal source and how to contact it. The application subsequently checks detailed freshness, format, and access conditions and retrieves data through the source's native interface.¶
A scheduler for an AI workload needs to find a compute service within its organisation or a cooperating provider. Discovery identifies the service and its contact information. Checking accelerator availability, jurisdictional constraints, and allocation policy can require subsequent application queries. The scenario in Section 5.2 is an instance of this example.¶
An AI application needs to find an inference service for a particular modality. Discovery supplies the relevant service type and communication information. Detailed model provenance, safety classification, and permitted use are assessed outside discovery where they exceed the minimum information.¶
A user or agent needs to discover a model-serving resource for inference. Discovery supplies its function and access method; subsequent interaction obtains richer model information when needed.¶
Discovery crosses organisational, administrative, or tenancy boundaries. A client may need to find a partner's resource, or a provider may expose different discovery views to its tenants, customers, or departments.¶
Organisations control what they publish and who can see it. Cooperation and authority to act for a user rely on existing arrangements. Discovery information may identify the publishing domain or responsible party to support later checks.¶
DAWN discovery needs to support discovery across administrative domains without requiring a single central registry. It also needs to support scoped discovery where the information returned depends on tenant, customer, department, network, or policy context.¶
Publication scope helps the client interpret a result. Discovery must also avoid exposing sensitive tenant or organisational structure to unauthorised parties.¶
An enterprise agent acting for an employee needs to discover an agent at a cooperating supplier. The discovery result indicates how to contact the supplier's agent and provides information needed by later authorisation and policy checks.¶
A software-as-a-service provider hosts agents or tools for multiple customers. Each customer needs discovery information scoped to its own tenancy. The same provider may also need different discovery views for internal operators and cooperating customers.¶
A research consortium operates AI services across several cooperating institutions with established trust relationships. Each institution publishes information about its own agents, data services, and tools. Collaborators need to discover entities by capability while respecting institutional boundaries and trust models.¶
A human operator, management system, or AI operations agent needs to locate AI resources and supporting tools for troubleshooting, audit, incident response, or compliance review.¶
This discovery category assumes that discovery is useful for both autonomous systems and human-operated tools. It also assumes that operational environments may be multi-vendor, multi-domain, and partly private.¶
Operational discovery needs to identify relevant resources and how to contact them. An operational client may subsequently obtain detailed runtime state, observability data, or compliance information. Those exchanges are outside discovery and may require separate permissions.¶
DAWN discovery may need to support tooling and operational inspection, not only agent-to-agent workflows. Discovery metadata may need to be logged, auditable, and understandable by operators.¶
Management and diagnostics for the discovery system itself may also be needed so that operators can understand discovery behaviour and failures.¶
Brokers, directory services, and operational tools may themselves be discovery targets when they support an AI system.¶
Observability of the discovery process is a separate operational consideration. Recording a discovery result or logging a later action does not itself constitute discovery of an operational resource.¶
A human operator needs to locate the AI agents maintained by a team within an organisation. Discovery provides their service types and communication information. Inspection of their runtime behaviour occurs subsequently.¶
An AI-assisted network operations system needs to discover telemetry sources, topology systems, control points, and remediation tools within an organisation or cooperating group. DAWN can help identify these supporting AI resources and their communication methods. Diagnosis, correlation, action recommendation, and remediation remain outside discovery. The managed-enterprise scenario in Section 5.4 illustrates how infrastructure-provided resource state and requester context can assist this bootstrap discovery.¶
An edge deployment includes lightweight agents that need fast and cacheable discovery because local connectivity is constrained. Discovery provides enough stable information for connection bootstrapping while avoiding reliance on rapidly changing operational state.¶
The taxonomy in [I-D.akhavain-moussa-dawn-problem-statement] provides background for classifying AI resources and their supporting entities.¶
The use cases in this document involve discovery information that may affect which entity is contacted, what protocol is used, and what trust indicators are presented. Incorrect or malicious discovery information could cause a discovering entity to contact the wrong entity, disclose information to an attacker, or use an unsuitable service.¶
Outside a local network, all discovery information exchanged between DAWN participants requires integrity protection. Local deployments may also provide such protection. Where it is absent, the deployment needs to account for the risks of unprotected discovery information; locality alone does not establish trust.¶
Discovery mechanisms also need to address authenticity, freshness, and access controls using established security mechanisms and trust arrangements. A cached digest or version comparison detects change but does not independently authenticate discovery content. This document does not select a protection mechanism or define new trust-evaluation methods.¶
Security of discovery metadata is distinct from trust in the discovered entity. Authentication, authorisation, attestation, and policy checks for later interaction are outside discovery.¶
Identity management and establishing trust relationships are also outside DAWN's initial scope; an authentic publication does not establish authority to act for a particular person.¶
Discovery queries may reveal information about the intent, capability needs, or operational state of the discovering entity. Published discovery information may also reveal information about deployed services, agents, data sources, models, or organisational relationships.¶
This suggests that a DAWN discovery mechanism will need to consider what information is public, what information is restricted, and what information should not be exposed through discovery.¶
Privacy-sensitive information can include proprietary capabilities, deployment location, capacity, runtime state, model or dataset metadata, requester identity, search history, and query intent.¶
Enterprise, cloud, and edge deployments have different needs for publication, caching, freshness, authorisation, observability, and operational control.¶
Dynamic information such as current load, availability, price, or status may change too frequently to be carried directly in a general discovery mechanism. Such application state can be obtained outside discovery through a referenced service. The minimum discovery information has its own freshness needs: an endpoint or supported protocol may also change.¶
A client may refresh retained information or repeat discovery when it is no longer sufficient. A changed description need not imply a different entity, and a temporary communication failure need not imply withdrawal. Recovery of application state or transfer of an existing task to a replacement is outside discovery.¶
Across administrative domains, operational considerations include publication authority, caching, freshness, visibility controls, logging, abuse handling, and failure behaviour.¶
This document has no IANA actions.¶
The authors thank Adrian Farrel and Jim Mozley for early discussion in the creation of this document, and their ongoing reviews and comments.¶
We also thank the following people for their DAWN mailing-list comments: Tommy Jensen (concrete use cases), Mohamed Boucadair (discovery starting points), Sumit Ahuja (capability discovery), Roman Danyliw (security and privacy), Matthew Quick (cross-domain applicability), Pete Resnick (agent identity considerations), and Mirja Kuehlewind (discovery steps).¶
Finally, we thank the following for their contribution to the use case practical examples:¶