Don't let GPT-6 Astra verify its own code; split builder, checker and reviewer
The author warns that Astra writes code, writes a test that agrees with it, then declares everything works, and suggests separate BUILDER, CHECKER and REVIEWER roles with evidence from failure paths.
this is f*cking gold if you’re building with GPT-6 Astra. it writes the code. writes a test that agrees with the code. then tells you everything works. that’s not independent verification. give the workflow three separate jobs: • BUILDER - implements the spec, without rewriting the acceptance criteria • CHECKER - runs independently defined tests, including failure cases • REVIEWER - compares the evidence against the original request for a checkout, “payment succeeded” isn’t enough. what happens when the callback arrives twice? when the refund fails? when the API times out after charging the customer? the happy path makes a convincing demo. the failure paths tell you whether you can ship. a second agent saying “looks good” is still not proof. require test results, observable behavior and unresolved failures. save this before your astra agent marks its own mistakes as done ↓