AgentGuard mines 642 real failure traces into guardrails and cuts abnormal execution from 69.0% to 26.7%.

Instead of hand-writing safety rules, AgentGuard generalizes recurring failure patterns from 642 documented coding-agent failures across 382 repository tasks into instruction-level constraints, activated conditionally so only the relevant rules enter the prompt. On a disjoint 100-task set with Claude Code on Haiku 4.5, successful completion went from 21.7% to 35.0%. The reusable part isn't their rule set, it's the direction of derivation: your own logged failures are the guardrail corpus, and conditional activation keeps the rules from bloating every turn.