How do agents advertise their capabilities in a multi-agent system?

Capability advertisement is how an AI agent describes the tasks it supports, the inputs it expects, and the interfaces other agents can use to request its services. These descriptions help a registry, orchestrator, or another agent select a compatible specialist without guessing what it can do.

Beginner-friendly example: API test-generation agents

Imagine a testing platform with three specialists:

• Requirements agent: Extracts endpoint rules from an OpenAPI specification.
• Security test agent: Generates authentication and authorization tests.
• Coverage reviewer: Checks whether generated test cases cover documented requirements.

The security test agent publishes a machine-readable description announcing that it supports generate_api_security_tests, accepts a validated endpoint specification, and returns structured test cases.

When an orchestrator needs tests for a payment endpoint, it can search for that capability, verify that the input schema matches, check whether the agent is authorized and available, and then send the job.

Describe capability → Publish trusted metadata → Discover a match → Validate contract and permissions → Invoke agent

What should a capability advertisement contain?

A useful description includes:

• Agent identity: Stable ID, owner, and version.
• Supported skills: Specific operations such as generate_api_security_tests—not a vague claim of “doing everything.”
• Input and output contracts: Schemas, required fields, supported formats, and relevant limits.
• Invocation information: An approved endpoint, protocol, task type, or internal routing identifier.
• Operational details: Supported environments, availability, or constraints when known.
• Security requirements: Authentication and authorization requirements, without exposing secret keys.

For example, an advertisement could look like this simplified JSON:

JSON
{
  "agent_id": "security-test-agent",
  "version": "1.0",
  "skills": ["generate_api_security_tests"],
  "input_schema": "EndpointRequirementsV1",
  "output_schema": "TestCaseListV1",
  "task_type": "security_tests.generate"
}

This is an illustrative internal schema, not a mandatory industry-standard format.

How can agents publish these descriptions?

• Agent registry: Register structured metadata in an organization-controlled catalog.
• Agent description document: Expose a discoverable capability document through an agreed protocol. Agent-to-agent protocols may define their own discovery documents or agent cards.
• Static configuration: Provide known specialist descriptions in code or deployment configuration.
• Service announcement: Publish approved metadata through a message bus or service-discovery layer.

The advertisement should be versioned and updated when a skill, schema, or invocation interface changes.

Advertising is not the same as proving ability

An advertised capability is a claim, not verified competence or permission. Before delegation, the system should check identity, contracts, available resources, policy, and—when important—measured quality.

Never treat an untrusted description as an instruction to grant broader tool access or execute an arbitrary URL or task.

LangGraph and Celery example

A LangGraph routing node can compare the requested capability with an allowlisted registry record and pass a scoped task to the matched specialist graph. For a Celery-backed deployment, an approved capability may map to a registered task type or queue. Both should validate the input schema and tenant permissions before execution.

In short

Capability advertisements make agents discoverable and interoperable by defining what they offer and how to request it. Clear contracts, versioning, and authorization turn those claims into dependable multi-agent handoffs.