Anthropic Confirms Three Bugs Quietly Degraded Claude Code for Weeks — and Promises a Stricter Rollout Process

Anthropic on Friday published a postmortem confirming what Claude Code users had been reporting for nearly two months: three separate bugs, all introduced via routine product changes rather than model retraining, quietly degraded the coding agent between March 4 and April 20. All three are fixed in version 2.1.116, released on April 20, and Anthropic has reset usage limits for affected subscribers as compensation. The company has also created a new public-facing X account, @ClaudeDevs, dedicated to product communication, and committed to a slate of process changes intended to prevent the same class of regressions in the future.

The Three Bugs, in Order

The first issue, deployed on March 4, lowered the default reasoning effort on Claude Code from "high" to "medium" in pursuit of faster latency. Internal evaluations did not flag the change as harmful, but users running real coding agents over real codebases noticed that Claude was missing edge cases and producing shallower fixes. The second issue, deployed March 26, was a caching bug: reasoning history that was supposed to be retained for an hour was being deleted after every turn instead, causing context loss within multi-step coding workflows and consuming user usage limits faster than expected. The third issue, deployed April 16, was a system instruction that capped Claude's responses at 25 words between tool calls and 100 words for final responses; it was intended to reduce verbosity but caused a measurable 3 percent quality drop on internal evaluations. None of the three issues affected the API directly — they were all artifacts of Claude Code's product layer.

The Process Fixes That Matter

Anthropic's stated remediation is more interesting than the postmortem itself. Going forward, more Anthropic employees will use the public Claude Code build instead of internal test versions — a dogfooding policy other AI labs have abandoned in favor of internal-only "champion" builds because they ship faster. All system-prompt changes will now have to clear broader evaluation suites before deployment. And compute-intensive changes will go through "soak periods and gradual rollouts" rather than instant global pushes. The implicit acknowledgement is that internal evaluations did not catch any of the three regressions because they did not measure the right things; the user complaints did. For Claude Code's professional user base — many of whom run agentic coding workflows that compound subtle quality drops into hours of wasted work — the postmortem matters because it concedes that "the API is fine" is not a sufficient quality bar when the product is the layer that actually faces developers.