Key ideas
How it works
An agent image implements a small contract so the platform can host it, check its health, and route requests to it:
Authentication is enforced by the platform — your agent code never validates tokens. The server listens on
AGENT_PORT (default 8080).
AGENT_PORT environment variable
The platform injects AGENT_PORT into your container at runtime. Your HTTP server must bind to this port — not a hardcoded value. The platform routes traffic and health checks to this port exclusively. If your server binds to a different port, health checks fail and the instance never reaches Running.
Manifest: /.well-known/agent.json
The manifest declares your agent’s identity and capabilities to the platform. It must be served as a static JSON response at GET /.well-known/agent.json.
Required fields:
If the manifest is missing or returns a non-200 status, the instance fails health checks and remains in
deploying state indefinitely. The platform does not retry — fix the image and redeploy.Register and deploy
1
Register the image
2
Deploy an instance
Common mistakes
- Confusing an agent with an instance. An agent is the registered image; an instance is a running deployment of it.
- Skipping the contract. Without
/healthand a manifest, an instance can’t pass health checks and stays indeploying.
Best practices
- Use clear, searchable capability labels.
- Version your images explicitly so deployments are reproducible.
Related pages
Compute → Agents
The full agent contract and registration.
Compute → Instances
Running deployments and endpoints.
Fleets
Group agents for orchestration.
Tools
Give agents external capabilities.

