ARTFEED — Contemporary Art Intelligence

Claude Code quality issues traced to harness bugs, not model

other · 2026-04-24

Anthropic's postmortem reveals that recent complaints about Claude Code's declining quality were caused by three separate bugs in its harness, not the underlying models. A notable bug on March 26 cleared Claude's older thinking from idle sessions but erroneously repeated every turn, causing forgetfulness and repetition. Simon Willison, who frequently uses long-running sessions, highlighted the impact. The postmortem underscores the complexity of debugging agentic systems beyond model non-determinism.

Key facts

  • High volume of complaints about Claude Code quality over past two months were grounded in real problems.
  • Three separate issues in the Claude Code harness caused material problems.
  • A bug on March 26 cleared Claude's older thinking from sessions idle over an hour, but repeated every turn instead of once.
  • The bug made Claude seem forgetful and repetitive.
  • Simon Willison reported having 11 stale Claude Code sessions at time of writing.
  • Willison estimates he spends more time prompting in stale sessions than fresh ones.
  • The postmortem was linked by Simon Willison on April 24, 2026.
  • The article is described as a link post by Simon Willison.

Entities

Institutions

  • Anthropic
  • Claude Code

Sources