AI team tools
Give AI employees tasks, follow each run, answer their questions, stop them, and see what is waiting for approval.
Updated 2026-10-13
These seven tools are the core of delegation. Results come back as { ok, summary, … , next } with hints and a polling interval. See Results and errors.
The AI team
list_team
See who is on the team and what each can do. Lists the AI employees in the connected workspace so you can choose who to give work to. Archived employees are left out.
Reads
No inputs.
Returns. The workspace id and a team list. Each employee has id, name, role, status, tools, connectors and autopilotFor (the apps they may act in without asking). With no employees the summary says Pytheus works alone.
- A
nexthint suggestsassign_task.
assign_task
Hand an employee a task. Returns a run id at once. Hands a task to one employee. It works on its own computer and keeps going after the call returns. Write a self-contained brief: the goal, who it is for and what done looks like.
Does work, uses credits
| Input | Type | Required | What it is |
|---|---|---|---|
employee | string | Yes | The employee’s name (from list_team), role or id. Matching ignores case. |
brief | string (8+ characters) | Yes | The task: the goal, who the result is for, and what done looks like. |
idempotency_key | string | No | Send the same key when retrying so the task starts once. |
focus | `auto` · `code` · `web` · `documents` · `research` | No | A tool focus for this task only. See Focus families. |
Returns. run with id, status and the employee; reused (true if the same key already started it); and poll_after_seconds (20). A next hint points to get_run.
- Uses Pythe credits and computer time, exactly as work started in the app does. Plan limits, credit guards and approvals all apply.
- Errors:
not_found(no such employee),ambiguous(more than one match: use the id),paused,employees(the plan has no employees),insufficient_credits. - The run is tagged with the client’s name, so the founder sees “via Claude” in Pythe.
list_runs
List recent and running tasks. Lists recent tasks the AI team has worked on or is working on, newest first.
Reads
| Input | Type | Required | What it is |
|---|---|---|---|
status | `active` · `waiting` · `completed` · `failed` · `any` | No | active is queued or running. waiting needs an answer or an approval. failed includes cancelled. Default any. |
limit | integer 1 to 30 | No | How many to return. Default 15. |
Returns. A runs list. Each has id, employee (Pytheus for the cofounder), goal (first 160 characters), status, progress (0 to 1), waitingFor, error, startedAt, finishedAt and via (which outside app started it).
get_run
Read status, report, files and waiting approvals. Reads one task: its status, the last steps, the finished report, the files and Library items it made, anything it is waiting to ask, and any approvals waiting for the founder.
Reads
| Input | Type | Required | What it is |
|---|---|---|---|
run_id | string | Yes | The run id from assign_task or list_runs. |
Returns. run with id, employee, goal, status, progress, report (the summary), verdict, files (last 10), library (last 10, each with an id and title), waitingFor, error, errorCode, recentSteps (last 8), computerMinutes, startedAt, finishedAt. Plus approvalsWaiting and poll_after_seconds (20, or null once finished).
- Poll until the status is
completed,failedorcancelled. nexthints say what to do:answer_runif it is waiting for input,list_approvalsif an action waits,get_library_itemto read a saved document.
answer_run
Answer a question an employee asked. Replies to a question a task is waiting on. get_run shows the question as waitingFor. The task carries on.
Does work, uses credits
| Input | Type | Required | What it is |
|---|---|---|---|
run_id | string | Yes | The run that asked. |
answer | string | Yes | The founder’s answer, in their words. |
Returns. The run id and its new status. A next hint points back to get_run.
- Only answer for the founder with what they actually told you. If you do not know, ask them first.
cancel_run
Stop a task. Stops a task that is running or waiting. Work already saved stays saved. Anything waiting for approval is closed.
Does work, uses credits
| Input | Type | Required | What it is |
|---|---|---|---|
run_id | string | Yes | The run to stop. |
Returns. The run id and its status (cancelled).
- Marked destructive in the tool’s annotations, so well-behaved clients ask before calling it.
Approvals
list_approvals
List emails, posts and pull requests waiting for approval. Lists what the AI team wants to do outside Pythe, such as send an email, post to Slack or open a pull request, and is waiting for the founder to approve. It is read-only. There is deliberately no tool that approves.
Reads
No inputs.
Returns. Up to 25 pending approvals, newest first. Each has id, runId, action, connector, askedAt and a preview of the first six payload fields (each clipped to 300 characters). Plus where_to_approve: the HQ address.
- Tell the founder what is waiting and where to approve it.
Related
Previous: Sign-in, workspaces and permissions · Next: Company and plan tools