Add appropriate human-checkpoint cadence
What you'll learn
Long Cowork tasks need scheduled human check-ins to catch drift. Cadence depends on length: hours → mid-task; days → daily; weeks → daily + weekly.
- Long-running tasks accumulate drift without checkpoints.
- Cadence is task-shape-specific: hours → mid-task; days → daily; weeks → daily + weekly.
- Checkpoint surfaces what's done, what's next, open assumptions, costly state.
Cowork tasks are long-running by design — multi-hour, multi-day, sometimes multi-week. The longer the run, the higher the cost of drift: the agent silently moves away from intent, accumulates wrong assumptions, and produces something useful-looking but off-target. **Human-checkpoint cadence** is the antidote: scheduled moments where the human reviews state, corrects course, and re-confirms intent. Without checkpoints, you have a runaway risk regardless of how good the model is.
The cadence is task-shape-specific, not one-size-fits-all. Short, well-scoped tasks (a few hours) may need only one mid-task check. Multi-day exploratory tasks need daily check-ins. Multi-week strategic work needs both daily progress reviews and weekly intent re-alignment. The decision is informed by *how fast intent can change* and *how expensive it is to discover drift late*.
What goes in a checkpoint: the agent surfaces (1) what it's done since the last checkpoint, (2) what it's about to do, (3) any open questions or assumptions it had to make, (4) any state that's surprising or costly to undo. The human reviews, corrects, re-confirms. The agent adjusts; the run continues. Good checkpoint design surfaces *intermediate artefacts* — partial output, written plans, code in progress — so review is concrete, not abstract.
The antipattern named in B9.3 ('set and forget') is the absence of cadence. Adding cadence is the structural fix; it's not 'micromanagement' — it's the discipline that keeps long-running agents aligned with the intent they were given hours or days ago.
Key points
- Long-running tasks accumulate drift without checkpoints.
- Cadence is task-shape-specific: hours → mid-task; days → daily; weeks → daily + weekly.
- Checkpoint surfaces what's done, what's next, open assumptions, costly state.
- Good checkpoints surface intermediate artefacts for concrete review.
- Cadence is the structural antidote to drift, not 'micromanagement'.
Examples
Daily checkpoint at start of day: 'Yesterday I gathered X. Today I plan Y. Open question: should source Z be excluded?' Human approves, redirects, or pauses. Run continues.
Daily progress checkpoint (concrete artefacts) + weekly intent re-alignment ('does this still match what we set out to do, given what we've learned?'). Two cadences, two purposes.
Pitfalls
- Setting one checkpoint at the end of a multi-day task. Drift is already cooked in.
- Making checkpoints abstract ('all good, continuing'). Without artefacts, the human can't see drift either.
- Treating 'no checkpoint' as a feature ('the agent is autonomous'). Cowork is explicitly *not* autonomous — it's collaborative.
Source notes: 00-academy-basics/notes/09-cowork.md
Ask Claude
Build a prompt with this lesson + your question, copy it, and open the Claude Project in a new tab.