Instrument every agent experiment to emit a cheap progress signal before final validation.

He et al. argue research agents run the same loop as a greybox fuzzer, and fuzzers make progress because coverage gives dense feedback on every execution that directs the next mutation. Scaling the proposer and ranking more samples post-hoc misses this entirely. Use the signal to pick the next intervention, not to sort finished runs, and keep your validation evidence protected from adaptive reuse.