Foundation 03 · Discovery and negotiation

Inspect the interface
before you delegate.

Capability discovery is a precondition for a legible handoff. Read what a peer declares, compare it with the request, identify unsupported operations and modalities, and keep authentication and authorization questions visible.

Learning boundary. Capability metadata describes an interface at a point in time. It does not prove that the capability works, that a peer is trustworthy, that a user has authorized the action, or that a declared skill produces a correct result.

Learning objective

Turn a capability card into a preflight decision.

By the end of this lesson, you should be able to compare required inputs and outputs with declared capabilities, modalities, extensions, authentication requirements, and unsupported operations before a workflow hands over context or authority.

01 · Surface

What can the peer expose?

List resources, tools, prompts, skills, endpoints, modalities, and extension points. Do not summarize a broad mission as if it were an invocable capability.

02 · Fit

Does the task match?

Compare required inputs, output types, latency, state, and failure handling with the declared interface. Record mismatches as a hold or escalation.

03 · Access

Who may use it?

Separate discovery from authentication and authorization. A visible endpoint or Agent Card does not grant permission to send data or request an effect.

04 · Change

What if the surface moves?

Record the version, extension, source, freshness, and recheck trigger. A capability declaration can become stale without the route disappearing.

Practice lab

Choose a peer without over-reading its card.

Two remote agents advertise “research assistance.” One accepts text and files with asynchronous tasks; the other accepts text only and lists a custom extension. Neither public description states whether external actions are permitted.

  1. Write the minimum capability set needed for the requested workflow.
  2. List what each declaration supports and what it leaves unknown.
  3. Define the authorization and data-sharing checks required before sending context.
  4. Specify the hold condition if the task needs a modality or effect that is not declared.

Source shelf

Study discovery models and application consent.

  1. Model Context Protocol specificationStudy capability negotiation and the server/client features for resources, prompts, tools, sampling, roots, and elicitation, alongside the specification’s application-level consent principles.
  2. Agent2Agent Protocol specificationRead the Agent Card, capabilities, skills, modalities, authentication requirements, extensions, and capability-validation patterns before designing a remote-agent handoff.