Case · 2026-04-23
Claude Code harness drama, models plead not guilty
The brain was fine. The body that types into your repo had opinions.
Summary: After users reported Claude Code quality had slipped, @ClaudeDevs published a postmortem on three issues in the Claude Code / Agent SDK harness (also impacting Cowork). Fixed in v2.1.116+. Models and Claude API said not regressed. Usage limits reset for subscribers.
Source: @ClaudeDevs · 23 Apr 2026
What landed
@ClaudeDevs wrote that over the past month some users reported Claude Code quality had slipped. They investigated and published a postmortem on three issues they found. All were fixed in v2.1.116+, and they reset usage limits for all subscribers.
A follow-up in the same thread stated the issues stemmed from the Claude Code and Agent SDK harness, which also impacted Cowork (runs on the SDK). Official line: the models themselves did not regress, and the Claude API was not affected. They described process changes (dogfooding configs matching users, broader evals).
Full write-up linked: anthropic.com/engineering/april-23-postmortem
Why it matters
Agent slop is often harness slop. Users experience “Claude got dumb” when the loop, tools, or session state go wrong even if the base model card is clean.
Satire margin
“The models didn’t regress” is the 2026 version of “it works on my machine,” except the machine is a multi-agent orchestra.
Primary sources
← Timeline· All cases· See a sibling failure? Submit it