Agent Class

The local Agent exposes these core public entry points:

• new Agent(options)
• agent.start(options?)
• agent.stop()
• agent.session(sessionId?)
• agent.sessions()
• agent.plugins

Constructor options

• id
• path
• tools
• instruction
• services
• plugins
• platform
• configureSession

Core semantics

• Agent itself is the local execution shell
• the real execution object is the session it creates

From an API-design perspective, it behaves more like a session factory plus a long-lived runtime entry point.

start() and stop()

agent.start() starts the long-lived runtime capabilities of the current agent.

Typical examples include:

• starting the HTTP server
• starting the local RPC server
• starting service lifecycles

Recommended usage:

await agent.start({
  http: {
    host: "127.0.0.1",
    port: 15314,
  },
  rpc: true,
});

If you do not call agent.start():

• you can still call await agent.session() and then session.prompt()
• but no HTTP or RPC endpoint starts
• and no background service lifecycle starts

agent.stop() closes whichever long-lived capabilities are currently running.

Lifecycle entry

But for most user scenarios, the recommended entry is:

await agent.start({
  http: {
    host: "127.0.0.1",
    port: 15314,
  },
  rpc: true,
});

The local Agent no longer exposes fine-grained transport properties like agent.http or agent.rpc.

If you need:

• HTTP only: agent.start({ http: { host, port } })
• RPC only: agent.start({ rpc: true })
• both: agent.start({ http: { host, port }, rpc: true })

agent.plugins is now an important public surface of the local SDK.

You can use it directly for:

• agent.plugins.list()
• agent.plugins.availability(pluginName)
• agent.plugins.runAction({ plugin, action, payload })
• agent.plugins.pipeline(pointName, value)
• agent.plugins.guard(pointName, value)
• agent.plugins.effect(pointName, value)
• agent.plugins.resolve(pointName, value)