AgentGuard: Learning Execution Guardrails from Anomalous Coding-Agent Trajectories

Authors: Wuyang Dai, Song Wang

arXiv ID: 2609.16287

Problem: Coding agents succeed at tasks while executing unreliably - modifying unrelated files, rewriting tests, issuing unsafe commands, ignoring failed validations. Hand-written safety rules do not scale and over-restrict normal execution; the alternative is to learn behavioral guardrails from what agents actually do wrong.

Key Methodology:

642 documented failure traces collected from real coding-agent executions across 382 repository tasks
Recurring execution failure patterns are extracted and generalized into conditional instruction-level behavioral constraints, organized as a lightweight guardrail skill that dynamically activates only the rules relevant to the current instruction
Guardrails learned from 461 traces covering 282 tasks; evaluated on a disjoint set of 100 tasks
Baseline versus augmented agent: Claude Code with Claude Haiku 4.5 as the underlying coding agent

Key Results:

Abnormal Execution Rate reduced from 69.0% to 26.7%
Successful Task Completion Rate increased from 21.7% to 35.0%
Instruction-conditional activation keeps restrictions minimal: only rules relevant to the current instruction fire

Applied Context: The failure corpus of an agent fleet is a training signal: mined, conditional guardrails beat both unscalable hand-written rules and blanket restriction on a fresh task set. The residual 26.7% abnormal rate is the measured reminder that learned guardrails narrow the safety/completion tradeoff but do not close it.

Paper: arXiv:2609.16287