I noticed a real problem while doing agentic coding in a plan → implement → test loop: the tests always pass.
The agent writes a huge number of tests, and they all succeed. The project always looks green. Why? Because the agent holds the full session history and every project detail. With all that context, it just tries to make every test pass so the project looks good — and the real errors stay buried in the code. The same happens even when I spawn a sub-agent.
The fix that worked for me: give the test agent minimal context.
When I spawn a testing sub-agent, I hand it almost nothing — just the goal of the current implementation. Then I ask it to do adversarial testing: try to falsify, attack edge cases, break the current implementation. A very short prompt that only says: falsify this implementation against the goal.
That instantly surfaced real errors. Fixing them made the project run smoothly.
The lesson, again, is that context matters in testing. Hand the agent the whole implementation history and it falls into the same trap — pretending to make things work. Hand it only the current aim and ask it to attack, and it does the real work.