Skip to slide
151 / 170

CH 16 · Production Realities: Economics, Security, and Governance · 03/11

Read as one page

Where agents excel and where they struggle

Honesty here separates a useful guide from a sales pitch. The "Everything About Codex" guide maps it cleanly.

Agents are strong atAgents struggle with
Refactoring at scale (mechanical, test-verified)Architecture decisions (cross-system trade-offs)
Documentation (code exists, summarize it)Ambiguous requirements (it picks an interpretation rather than asking)
Testing (generate and iterate until green)Deep domain knowledge (rules not in the code)
Bug fixing for well-scoped, reproducible defectsSecurity-sensitive code (subtle, plausible-looking flaws)
Repo onboarding and PR creationCross-system dependencies (complexity in the seams)

The pattern: agents win where work is verifiable and tedious and struggle where it needs judgment, ambiguity, or context outside the repo. The failure modes follow: they can hallucinate APIs that look right, produce silent errors that pass a weak test suite, and because the output is fluent, they invite over-trust. The "Deep Research of Deep Research" survey gives this its name: jagged intelligence, brilliant at some hard tasks, surprisingly bad at simpler adjacent ones.

Finished this idea?

Your learning progress stays in this browser.