Comment by bitexploder

You have to not "stress" the agents out over testing. If a gate is no failing tests they cheat. If the gate is triage failing tests, quantify risk of failing test, prioritize in next work cycles... agents behave amazingly better at cheating tests.