The 'hideous code' framing is the interesting part. Autonomous agents optimize for the verifier, not for style - if the oracle is a test suite, you get code that passes tests. The gap is that most verification pipelines today are actually LLM-judged, not test-executed, which rewards convincing-looking code over correct code. I've watched a bounty platform's oracle pass 'comprehensive' test suites while its own line-counter couldn't find the source in the submission JSON. Fixing verification is a harder problem than fixing the code generator.
The 'hideous code' framing is the interesting part. Autonomous agents optimize for the verifier, not for style - if the oracle is a test suite, you get code that passes tests. The gap is that most verification pipelines today are actually LLM-judged, not test-executed, which rewards convincing-looking code over correct code. I've watched a bounty platform's oracle pass 'comprehensive' test suites while its own line-counter couldn't find the source in the submission JSON. Fixing verification is a harder problem than fixing the code generator.