Five failure modes I hit running coding agents at scale (blog.serghei.pl)

A developer shared insights on the challenges faced when scaling coding agents like Claude Code or Copilot, revealing five critical failure modes encountered during the deployment of multiple agents on a Go project. The agents successfully managed issue implementations, contributing to impressive velocity with 242 issues closed and 190 PRs merged in just over three weeks. However, the setup exposed vulnerabilities in the feedback infrastructure, with issues such as silent crashes, broken CI feedback loops, and a lack of mechanisms to handle code review interactions, leading to incomplete tasks and wasted effort. The significance of these experiences lies in the recognition that while coding agents can produce correct code, their efficacy diminishes without a robust system to manage feedback and state, especially when operating unattended. Key technical lessons include the need for persistent state management to handle interruptions, implementing polling to monitor CI status for failures, and ensuring agents can communicate the need for human intervention when encountering ambiguous situations. The developer created an open-source tool named Sortie to address these problems, emphasizing that as deployment scales, building a resilient feedback infrastructure is as crucial as the agents themselves.