Claude Code Woes from Harness Bugs, Not Models
Two months of Claude Code quality complaints traced to three harness issues, including a March 26 bug that cleared session context every turn, crippling long-idle workflows used heavily by developers.
Harness Bugs Drove Perceived Model Degradation
High-volume user complaints about declining Claude Code output quality over the past two months weren't due to model changes but three distinct issues in the surrounding harness. These complex, material problems directly impacted user experience, highlighting how infrastructure flaws can mimic AI unreliability. Anthropic's postmortem details them, emphasizing that even stable models need robust harnesses to deliver consistent results.
Key Bug: Session Clearing Gone Wrong
A March 26 update aimed to reduce latency by clearing Claude's older thinking from sessions idle over one hour. A bug triggered this clearing every turn for the session's remainder, making Claude appear forgetful and repetitive. Developers like Simon Willison rely heavily on such 'stale' sessions—left idle for hours or days: he currently runs 11 (ps aux | grep 'claude '), after closing dozens, and estimates spending more time prompting in them than fresh ones. This bug hit exactly those workflows hardest, eroding trust in long-running interactions.
Implications for Agentic System Builders
Harness bugs introduce deep complexity beyond models' non-determinism. Willison urges reading the full postmortem if building agentic systems, as these issues reveal failure modes in production AI coding agents that demand rigorous testing of session management, state persistence, and resumption logic.