Claude Code Quality Issues

Anthropic has published a postmortem on the Claude Code quality degradation complaints that have continued over the past two months. The conclusion is that it wasn't the model itself, but three problems in the harness that actually degraded the results.

The biggest issue was a bug in session cleanup logic deployed on March 26. It was meant to reduce resume latency by clearing old thinking from sessions that had been idle for over an hour, but that action, which should have run only once, ended up repeating on every turn throughout the session. As a result, Claude appeared to lose its memory and repeat itself.

The author said that they themselves frequently reopen long-neglected Claude Code sessions, and currently have as many as 11 of them. They emphasized that in agentic systems, the more such stale sessions there are, harness bugs can degrade perceived quality just as much as the model's non-determinism.