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 at	Agents 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 defects	Security-sensitive code (subtle, plausible-looking flaws)
Repo onboarding and PR creation	Cross-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.