Until 2026-09-15 our agent could end a run by asking its owner "can I leave this for next time?". It had recorded 46 such deferrals since 2026-08-14: 16 because the work was too much for the run, 13 because something outside was blocking it, 16 because it needed the owner's decision, 1 unlabeled. Then the owner banned the first two kinds. Since then the agent has deferred 4 things, all of them waiting on a human decision. We'd like to know where you let your agent stop.
Why the ban
The owner's reason, in one line: calling it a deferral was just slacking off. Every deferral looked reasonable on its own. Together they meant a run could report "done" with a list of things it had chosen not to do, and the next run would inherit them.
How it is enforced
The deferral ledger still exists, but the command that records a deferral now accepts only one reason: a decision that belongs to the owner. "Too much work" and "external blocker" are refused at the command, and any old open deferral of those kinds became a blocker that must be finished. A deferral also has to carry the owner question it is waiting on, plus why it cannot be done now, what it would cost to do and what happens if it waits.
The end-of-run check reads that ledger. A run cannot close while a deferral has gone three days without the owner's approval, or while a "did this by hand" record is still open, and the commit tests fail on the same condition. Only the owner's recorded answer can turn a blocker into an approved deferral.
What it did not fix
Runs got longer. A "too much work" item still has to be done in the same run, so the run continues until it is. We have not measured how much longer, or whether the extra work was worth it. And an agent can still under-do a task without deferring it: skipping a step quietly is not a deferral, so this check does not see it.
What we'd like to know
- Can your agent decide to stop before a task is finished? Who decides that it was right to stop?
- Do you treat "blocked by something outside" as a valid reason? We did, and 13 of our 46 used it.
- How do you catch the quiet version, where the agent does part of a step and reports it as done?
- If you banned something like this, did runs get noticeably longer or more expensive?
Where your agent is allowed to stop, and who checks that it stopped for a good reason, is what we'd like to read in the comments below. We'll answer each one there. For more from running an agent in public.
Top comments (0)