Verifying Stop Scope in Agent Workflows

This study defines an interesting problem, along with a useful distinction between stopping an agent process and its sub-parts: cancelling a run, waiting for a worker to drain, blocking a pending write, and stopping computation are all different, and the author's Open-SWE experiment makes that concrete with real file effects. I liked the honesty throughout - the author is careful that two fixed seeds can't establish a rate, that failing V wasn't a broken guarantee, and that the 286-case catalogue is a proposal, not results. Two things hold it back. First, the empirical base could be increased from 26 conditions on 2 seeds, and run on multiple hosts. Second, and most importantly, the paper could be made easier to read - the introduction especially is long and reflective, and the actual research questions would land better if they came first. Overall, strong execution - the verifiable evidence attached is rare and commendable.

The conceptual contribution is the useful part. Decomposing what it means to stop an agent into three separately checkable requirements — no final report written, no write from revoked authority, no lingering effect once completion is acknowledged — gives an incident responder something concrete where there was one ambiguous word.

The evidence is also unusually reproducible. The released bundle matches its published size and hash, the project's own auditor returns every headline figure as stated, and the audit records back the results table on inspection. That standard is rare.

What it does not do is test the thing the paper is about. Every agent response was scripted and fully compliant. The introduction spends its opening pages on agents that might still be working after a stop, and on whether child agents can still be spawned once one is issued — and the experiment then gives an agent no opportunity to do either. What was validated is the control layer against a cooperative actor, which is the easy case. A few runs against a small open-weight model would have put one real behavioural condition under test.

Two practical things. The linked repository holds little more than a pointer to an external archive, so anyone following it finds a signpost rather than the work; moving the harness, the auditor and the summary records into the repository itself is a cheap fix that would let the rigor here actually be seen. And the opening section buries the contribution behind a long account of how the author arrived at it — leading with the three requirements, and putting the reasoning after them, would serve the ideas better than the current ordering does.