Skip to content
tatu

Subagents

A subagent takes a specific task, works in its own conversation and reports back to the model. This keeps searches, file contents and dead ends out of your main conversation.

For example, ask:

Use subagents to investigate the parser and the tests separately, then summarize what needs to change.

The model decides when to call the agent tool. It can delegate independent tasks together; each subagent gets the task it is given, not your whole conversation.

Types and model tiers

Section titled “Types and model tiers”

Two types come built in:

  • general-purpose investigates questions and carries out multi-step work. It uses balanced by default.
  • explore is instructed to search and read without changing files. It uses fast by default.

The model can choose a tier for a task: fast for searching and summarizing, balanced for well-specified implementation, or main for work needing more judgment. main uses your session model. The other tiers use the catalog’s choice for your current provider, falling back to your session model when needed.

Use the balanced and fast tabs in /model to choose their models and effort. For your own types, see Custom subagents.

Look inside

Section titled “Look inside”

Press Alt+A to inspect subagent conversations. Their final reports go back to the main model, which brings the relevant results into your conversation.

In the Tatu app, tap a subagent card to follow its full conversation live. Expand Task for the instructions and Details for model, tokens and cost; tap tools to inspect their output. The header switches between subagents. Scrolling up stops following new activity; Latest resumes it. Back returns to your main conversation without losing your place or draft.

This screen is read-only: send messages and answer questions in the main conversation. After a connection drops, it refreshes before following again. An older computer version or unavailable history shows an explicit notice with the task and report that are available. The remote terminal viewer is not integrated yet.

Subagents cannot ask you questions or start more subagents. Their usage counts toward the session total; /cost shows it separately by model.

See also

Section titled “See also”