Anthropic identifies three bugs causing Claude Code quality degradation over two months

Anthropic has published a postmortem confirming that user complaints about Claude Code quality issues over the past two months were caused by bugs in the coding assistant's infrastructure, not the underlying AI models.

According to the company's analysis, three separate issues in the Claude Code harness created "complex but material problems which directly affected users."

Critical session bug affected idle workflows

The most significant bug stemmed from a March 26 update intended to improve performance. According to Anthropic: "We shipped a change to clear Claude's older thinking from sessions that had been idle for over an hour, to reduce latency when users resumed those sessions. A bug caused this to keep happening every turn for the rest of the session instead of just once, which made Claude seem forgetful and repetitive."

This bug particularly impacted developers who maintain long-running sessions. Simon Willison, who flagged the postmortem, noted he frequently leaves Claude Code sessions idle for an hour or longer and currently maintains 11 such sessions simultaneously. He estimates spending more time in these "stale" sessions than in recently started ones.

Harness complexity revealed

The incident highlights the technical challenges of building reliable AI coding assistants beyond model performance. All three identified bugs occurred in the harness layer that connects users to the underlying Claude models, rather than in the models themselves.

Anthropic has not disclosed specific details about the other two bugs or when fixes were deployed. The company also has not provided information about compensation or service credits for affected users during the two-month period.

What this means

This case demonstrates that AI coding assistant quality depends on more than model capabilities alone. The infrastructure managing context, session state, and user interactions introduces its own failure modes that can significantly degrade user experience. For teams building agentic systems, these "harness bugs" represent a distinct category of technical debt separate from model improvements. The fact that these issues persisted for two months before public acknowledgment also raises questions about monitoring and quality assurance processes for production AI systems.