AgentGuard learns execution guardrails from 642 real failure traces and cuts Claude Code's abnormal execution rate from 69.0% to 26.7%

arXiv 2609.16287·high signalXBlueskyLinkedInCopy link

Rather than hand-writing safety rules, AgentGuard mines recurring failure patterns from 642 documented coding-agent failure traces across 382 repository tasks, generalizes them into instruction-level constraints, and packages them as a guardrail skill that activates only the rules relevant to the current instruction. Evaluated on a disjoint 100-task set with Claude Code on Haiku 4.5, it drops the Abnormal Execution Rate from 69.0% to 26.7% and raises Successful Task Completion from 21.7% to 35.0%. The reusable technique for anyone running agents in a repo is the direction of derivation: your own logged failures become the guardrail corpus, and conditional activation keeps the rules from bloating every prompt.