Tracked tasks
A tracked task has a brief, a definition of done, its own channel and a progress log. Its work is checked against the definition, sent back to improve, then handed to you.
Updated 2026-10-13
A tracked task is how you hand over work you care about getting right. It is the right tool for anything with a clear deliverable.
Hand it off
Open Hand it off from a plan task, from the composer’s + menu, or from the command palette. “It gets its own channel, a progress log and a check against what done looks like.”
| Field | What it does |
|---|---|
| Who does it | Pytheus (AI cofounder), one of your employees, or Me, in which case Pytheus checks in on you. |
| Task | A short title, up to 200 characters. |
| Brief | Context, constraints, links and what you already know, up to 6,000 characters. |
| Done when | One checkable line each, such as “Covers 5 direct competitors”. Leave it empty and the check uses the brief. |
| Budget (credits) | The most this task may spend. It stops and tells you when it gets there. Reviews, check-ins and every step count. Leave it empty for the plan default. |
| Send it back to improve when a check fails | On by default for agent-owned work. |
| Check in | For work you keep: Never, twice a day, daily, every 3 days or weekly. |
The loop
- Hand-off. The task row and its channel are created, the brief is posted in the channel, and work starts. The plan task moves to in progress.
- Work. The owner gets a run in the channel. Each pass is its own run.
- Check. When the run ends, the work is scored against the definition of done, using the real files and the deterministic checks.
- Improve. If the work is short and review rounds remain, it goes back to the owner with notes. The default is two rounds; the range is zero to four.
- Ready. When it passes, or the rounds are used up, it goes to you with an honest statement: either that the checks passed, or that some of it is still short and it is yours to decide.
- You decide. Approve it, or request changes.
Statuses
| Status | Meaning |
|---|---|
| Open | Created, not started. |
| Working | The owner is on it. |
| Checking | The work is being scored against the definition of done. |
| Revising | Sent back to improve. |
| Ready | Waiting for your decision. |
| Done | You approved it. |
| Blocked | Stopped, with the reason, such as reaching its budget. |
| Cancelled | Stopped by you. |
The review
A review has a verdict (pass or needs changes), a score, a summary, each criterion marked met or not, improvements, and the evidence measured. A criterion is marked as counted when code measured it, or model when the reviewer judged it. You can also ask “check my work on X” in any chat: the owner checks it in the right channel, and anyone else points you to the owner’s channel.
Budgets
- Every model call that works on a task is recorded against it, so its spend is a sum, not an estimate.
- With no explicit budget, a task gets a default of 15% of your plan’s monthly credits, never under 5 and never over 50 credits (20 when the plan has no credit limit).
- At the budget the task stops and says so, with the numbers: “This task reached its budget: 6.2 of 5.0 credits spent.” You raise the budget to continue, or approve what exists.
- A worker sets aside an estimate before each model call, so several workers starting together cannot all pass the same stale balance check.
Work you keep
If you assign a task to yourself, Pytheus introduces itself as the checker and schedules check-ins at the cadence you chose. You can post your work for a check, and the same review runs.