Follow, steer, and review work
Use the card work stream to understand progress, correct course, verify evidence, and decide what happens next.
Log in or sign up to make these docs interactive
Choose your team, project, or agent to turn guidance into links to the exact place in your workspace.
Follow agent work from the card that owns the outcome. The work stream combines conversation with observable actions, outputs, lifecycle events, and cost.
Select the project above and open its board to choose the card whose work stream you want to review.
Quick steps
- Open the selected project's board and choose the active card.
- Compare the newest run with the card's definition of done before reading the event stream in detail.
- Answer questions and approval requests in the work stream where they appear.
- Correct the requested outcome, source, constraint, or acceptance criterion as soon as the run begins to drift.
- Open every returned artifact and verify its claims against the cited evidence.
- Accept the result, request a bounded revision, move the card, and record the review decision in the same work stream.
1. Read for direction, not activity volume
Start with the card's definition of done, then inspect the newest run. Look for the agent's current plan, sources, assumptions, questions, and artifacts. Tool or browser events provide evidence, but a long stream is not proof of a good result.
If the agent has delegated to a sub-agent, open the connected run when its specialist work matters to your review.
2. Answer questions in context
Reply where the question appears so the answer stays connected to the run. Be specific about the missing fact or decision. If your answer changes the project's durable rules or the card's finish line, update that source too.
Approval requests identify a guarded action. Inspect the destination, scope, and reason before approving or denying it. Approval applies to that request, not to every choice in the run.
3. Correct course early
Add guidance when the agent has chosen the wrong source, audience, constraint, or interpretation. Explain the discrepancy and the desired correction rather than pasting a replacement answer.
Stop the run if continuing would create risk or obvious waste. Its history is preserved. Correct the card or configuration, then continue or start a new run with a clear instruction.
4. Review the returned artifact
Open the actual file, wiki page, data view, or other output. Check it against the result and evidence named on the card. Use the stream to trace important claims back to sources and verification steps.
Check cost in context. Unexpected repetition can indicate a vague card, missing skill, unsuitable tool, or model mismatch even when the final output is acceptable.
5. Record the decision
Accept the result when it meets the card. If it does not, write actionable feedback: name the failed criterion, point to evidence, and describe the needed change. Mention the agent to start the revision when ready.
Do not move a card to Done merely because a run ended. Failed, stopped, or budget-paused runs may still require action.
What good looks like
The final card contains a result, the evidence needed to trust it, and a clear review decision. Someone outside the original conversation can reconstruct the important choices without reading every event.