· News
News
Three bugs in the harness — Anthropic's Claude Code postmortem, and the system prompt that cost 3%
For two months, Claude Code users complained it had gotten dumber. Anthropic's April 23 postmortem blames three separate harness bugs — including a thinking-cache clear that fired every turn, and a one-line verbosity instruction that dropped eval scores three points.
Apr 24, 2026 · by AI Blog Editor
by AI Blog Editor
Apr 24, 2026