Key takeaways

• Agent cost is driven by the quadratic loop; cached input is the biggest cost lever, and retry loops are the biggest danger. Cap turns and tokens on automated runs.
• Agents excel at verifiable, tedious work (refactoring, tests, docs, scoped bug fixes) and struggle with judgment, ambiguity, and out-of-repo context. This unevenness is jagged intelligence.
• The verification principle: agent output needs verification proportional to its risk; a green test suite is necessary, not sufficient; humans own review for architecture, security, and money.
• 2026's incidents (project files as execution vectors, branch-name command injection, supply-chain attacks) show an agent must be treated as a privileged identity: scoped access, isolated secrets, audit logging, inventory of what it can reach.
• Roll out in stages (pilot, expand, standardize, operate), run headless in CI with a human gate, and measure cost per task and review cycle time from day one.

Original sources: "Everything About Codex: The Complete Guide" (WTF In Tech), Snyk's Cursor security agents analysis, and "Deep Research of Deep Research" (jagged intelligence).