What is an agent registry in a multi-agent system?

An agent registry is a trusted catalog that records which agents are available in a multi-agent system, their capabilities, and the approved ways to invoke them. It is the directory behind agent discovery: discovery is the act of searching for suitable agents, while the registry stores the records that make those searches possible.

Beginner-friendly example: API test-generation platform

Imagine a platform with several specialized agents. The registry contains entries like:

• requirements-extractor: Reads API specifications; accepts a document reference.
• security-test-generator: Creates authentication and authorization test cases; accepts structured endpoint requirements.
• coverage-reviewer: Checks test cases against a requirement set; accepts a test-suite reference.

When an orchestrator needs security test cases, it queries the registry for the required capability and receives eligible agent records. It then checks permissions, current health, and input compatibility before invoking a specialist.

Register agent → Validate metadata → Discover suitable agents → Authorize invocation → Monitor or retire registration

The registry identifies an agent; it does not guarantee that an agent's advertised capabilities are correct or that the task will succeed.

What information should each registry entry hold?

A useful agent record includes:

• Identity: Stable agent ID, version, owner, and tenant or organizational scope.
• Capabilities: Clearly named skills or supported task types.
• Contract: Expected input/output schemas and supported protocols.
• Invocation details: Approved service endpoint, graph node, or queue identifier.
• Security metadata: Authentication method, required permissions, and allowed callers.
• Operational metadata: Health, capacity, version status, and last verified time.

For instance, the security-test-generator might advertise the capability generate_api_security_tests and require a validated endpoint-specification object as input.

How does the registry stay reliable?

1. Registration: An authorized deployment or administrator creates an entry.
2. Validation: Check identity, schema compatibility, and permitted capabilities.
3. Health and freshness: Update status using trusted checks or expiring registrations.
4. Version management: Support staged rollout and remove incompatible versions.
5. De-registration: Revoke or retire entries for stopped, compromised, or obsolete agents.
6. Audit: Record who added or changed a record and which version was used for each job.

Registry updates should be authenticated. Agents must not gain powerful tool permissions merely by claiming a capability in their metadata.

Agent registry vs. service registry

A service registry typically tracks service instances and network locations. An agent registry adds semantic metadata useful for agent selection, such as capabilities, task contracts, and delegation policies. A system may combine both, but the concepts are not identical.

Practical LangGraph and Celery design

In a LangGraph workflow, an orchestration node can query the registry for approved specialist types and route work only to defined nodes or subgraphs. For Celery workers, an entry may reference an approved task name or queue.

A relational database such as PostgreSQL can hold registry records; a cache may accelerate reads, but authorization and version checks should still rely on trusted, current data. Resolve candidate records to an allowlist, not arbitrary URLs or executable task names supplied by an LLM.

In short

An agent registry is the controlled source of information about agents, their capabilities, contracts, and permitted invocation paths. Good registry design makes dynamic collaboration discoverable, auditable, and safer as the number of agents grows.