Skip to main content
Agents are authored and published in the dashboard — see Your first agent. Over the public API you only ever name one, and read a small embedded form of it.

​
The id

agent_ plus 32 hex characters. It names one exact published version — an exact lookup, scoped to your workspace, with no latest-resolution. Publishing a new version mints a new id. Pass it when creating a thread:
A published agent also carries an mcp_url built from the same id — the address a user’s own AI client connects to. See MCP export.
The URL exists for every published version, but the door only opens once the agent’s connection has a login bridge.

​
As embedded in a thread

​
string
The agent’s name, as authored.

​
As embedded in a trace

GET /v1/loom/runs/:id/trace returns agent_snapshot — the complete spec that was in force when the run started, including instructions, model class, and wired patches with their permissions. Editing an agent does not rewrite history: an old trace shows the agent as it ran, not as it stands now.

​
Errors