Agent Hosts
Agent hosts are the LLM boundary. They should be used only when deterministic routing, guards, templates, or host calls cannot answer the question directly.
| Handler | Use it for | Reference |
|---|---|---|
host.agent.extract | Tiered structured extraction: synonyms, slot templates, then LLM. | ../hosts.md#hostagentextract |
host.agent.ask | Read-only inspection that returns prose and optional typed JSON. | ../hosts.md#hostagentask |
host.agent.decide | Schema-bounded verdicts and gates. | ../hosts.md#hostagentdecide |
host.agent.codeact | Bounded agent loop that emits capability-scoped Starlark snippets, then done(payload). | ../hosts.md#hostagentcodeact |
host.agent.task | Focused agent work with replay artifacts and acceptance modes. | ../hosts.md#hostagenttask |
host.agent.converse | Free-form conversational sessions with permission controls. | ../hosts.md#hostagentconverse |
For declaring alternate agent transports and providers, see ../agent-plugin.md, ../agent-providers.md, and ../agent-backends.md.
For the shared Starlark sandbox used by host.agent.codeact, see ../starlark.md.