Worker agents
Let the model hand parts of a task to faster worker models that run at the same time and report back.
2 min readUpdated for Chat 0.5.2Edit on GitHub
Ask for agents
Ask for the work in any chat: "explore this codebase with four agents and tell me how auth works", or "use sub-agents to compare these three approaches". The model splits the work into tasks and sends each to a worker with its delegate tool. The workers run at the same time on a smaller, faster model, and their reports come back to the model, which does the deciding and writing.
A slow, strong model plus several fast workers finishes sooner, and the main model's context stays small because it reads reports instead of every file.
Workers run inside Chat. They never use your connections or clusters, so asking for agents never starts Slurm or Kubernetes jobs.
What workers can read
A worker is a short chat of its own. It gets only its task, not your conversation, so the model writes each task to stand on its own.
- In a chat with a workspace, workers list, read and search its files and read its git history.
- When the task needs a repository or the project's code and the chat has no workspace yet, Chat turns the chat's workspace on first. Given a repository address, it clones the repository into
repos/and tells each worker where it is. From your next message the model can use the workspace itself too. - Otherwise, workers answer from what their model knows.
Workers can't change files, run commands or send work on to other workers.
Follow the work
The reply shows one step for the workers, with each worker's title, what it's doing now, its steps and tokens. Open a finished worker to read its report. A worker that fails says why, and the model works with the reports that came back.
Limits and usage
| Limit | Value |
|---|---|
| Workers per call | 6 |
| Steps per worker | 30 |
| Time per worker | 5 minutes |
| Output per worker step | 8,192 tokens |
| Workers running at once on one Chat replica | 24 |
Each worker is charged to you on its own usage record, under the worker model, so your daily allowance and usage by model include it. Workers use low thinking effort.
Administrators choose which models can be workers; see Models, groups and limits.