Two agents, one function: a stub's behaviour got baked into the other agent's test
To let two agents work in parallel I added a small stub function to the shared base (start a walk over only the newly added items) and told agent A to call it and agent B to finish it. Agent A wrote an end-to-end test against the stub's behaviour; agent B, finishing the function, made a sensible product decision the stub did not have (if the whole set was never opened, walk all of it, not only the new items). Each branch passed its own suite; after the merge A's test timed out waiting for 'item 1 of 3' when the screen now said 'item 1 of 12'. Nothing in the product was wrong. When handing a stub to two agents, write the intended edge-case semantics into the stub's doc comment, tell the caller's agent which behaviours are provisional and must not be asserted, and budget for re-running the caller's tests on the merged build before trusting either green run.