Detect set-and-forget antipattern
What you'll learn
Set-and-forget is dispatching a long-running Cowork task and not checking on it. It's an antipattern — the right shape is set-and-steer with checkpoints.
- Set-and-forget = dispatch + ignore until end. Antipattern by design.
- Detection: multi-day, zero mid-task interaction, no intermediate artefacts, 'autonomous' framing.
- Fix: add human-checkpoint cadence (B9.2). Transform to 'set and steer'.
**Set-and-forget** is the antipattern of dispatching a long-running Cowork task and not checking on it until the end. It's tempting because Cowork's persistence makes it possible — the task just keeps running. It's the wrong default because long-running agents drift; the longer the run, the larger the drift; without intermediate review, you discover the drift only when it's expensive to undo.
Detection signature: (1) the task spans multiple days with zero mid-task interaction; (2) no intermediate artefacts are produced (or none are reviewed); (3) the only feedback loop is end-of-task; (4) the operator describes the task as 'autonomous' or 'set and forget' (Cowork is explicitly *not* autonomous; the design is collaborative). Any of these is the same antipattern.
The structural fix is human-checkpoint cadence (B9.2). Adding cadence transforms set-and-forget into 'set and steer' — the intended Cowork shape. The agent does the heavy lifting; the human's lever is intermediate review and course correction. Both are first-class roles.
Adjacent failure: producing no intermediate artefact during a long-running task even if the cadence is set. The cadence is the *opportunity* for review; without artefacts to look at, the review is empty. Fix the artefact-emission shape (write the partial output to a known path, log the in-progress plan, snapshot the code state) so the cadence has something concrete to inspect.
Key points
- Set-and-forget = dispatch + ignore until end. Antipattern by design.
- Detection: multi-day, zero mid-task interaction, no intermediate artefacts, 'autonomous' framing.
- Fix: add human-checkpoint cadence (B9.2). Transform to 'set and steer'.
- Adjacent failure: cadence without artefacts to review = empty cadence.
Examples
Operator says: 'Just run this 4-day research task and ping me when done.' No intermediate artefacts; no checkpoints; no opportunity to correct drift. Day 4 reveals the agent went off-track on day 1 and built on the wrong foundation.
Same task, with daily checkpoints surfacing the day's findings, the next day's plan, and any open questions. Day 1 drift is caught on day 2 morning; correction is cheap; the rest of the task stays aligned.
Pitfalls
- Treating Cowork's persistence as 'autonomy'. It's not; it's a managed long-running surface that *expects* steering.
- Setting cadences but providing nothing concrete to review. Empty cadence = false confidence.
- Disabling cadence after a few smooth runs. Drift compounds; missing cadence is hard to put back.
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.