Figma's security team describes how they built an agentic security system that guards code at three stages: generation, pull request review, and historical auditing. The system uses Claude Code (Opus 4.8) and Codex (GPT-5.6) to review every PR at ~$0.50 median cost, with a 70%+ precision target before enabling developer-facing comments. A key insight is that precision must be improved before recall, and that precision and recall require different measurement approaches — precision from live PR author feedback, recall from replaying against known-bad commits. The team built a 99-line, 2,560-word policy of 68 precedents that doubles as a threat model, shared across all three agent stages. Self-improvement loops automatically update the policy when new bugs are found or false positives are reported. An adjudicator pass raised recall by ~30%. Repo-wide auditing found 100+ latent vulnerabilities including two criticals missed by SAST. Agent hooks during code generation steer toward secure patterns just-in-time, reducing certain bug classes by ~50%.
