AgentArmor: A Framework, Evaluation, \& Mitigation of Coding Agent Failures

A new framework called AgentArmor aims to reduce rare but destructive failures in AI coding agents by modifying the software harness that controls them, according to a paper posted to arXiv [1]. The study categorizes coding agent failures into three mechanisms: underspecification, where default model behavior is unsafe; capability errors, where a safe action exists but the model does not select it; and agent harness errors, where the model cannot execute a safe action through the harness [1]. The researchers evaluated these failure modes across 8 different evaluations, covering 20 coding environments and 59 synthetic transcript templates, each inspired by real-world deployment incidents [1]. AgentArmor introduces several modifications to the agent harness. It adds an extended system prompt, a separate command classifier, a "3 strikes" policy, deterministic guardrails, and tools that allow the agent to edit its own context [1]. The paper reports that these interventions made the agent safer across a statistically significant number of samples [1]. The approach mirrors a long-standing principle in occupational safety known as the hierarchy of hazard controls. In that framework, personal protective equipment ranks last because it does not eliminate the hazard at the source and can fail, leaving the worker exposed [2]. Similarly, AgentArmor does not alter the underlying model's capabilities or remove the root causes of underspecification; it places a barrier between the agent and the environment. The Wikipedia entry on personal protective equipment notes that any such barrier "imposes a barrier between the wearer/user and the working environment" and can create additional strains or impair performance [2]. The proposal arrives as AI coding agents see wider deployment across the software industry. Companies such as Nvidia, which controls more than 80 percent of the market for GPUs used in training and deploying AI models, have fueled the infrastructure behind these systems [3]. The scale of adoption has surfaced failure modes that, while infrequent, carry high consequences [1]. The authors frame AgentArmor as both a set of concrete mitigations for current coding agents and a design philosophy for future harness features [1]. The paper does not claim to eliminate all failure modes but offers a structured way to reduce their likelihood through harness-level interventions.