Claude Code Degradation, Caused by Bugs

In an April 23 postmortem, Anthropic explained that the quality degradation in Claude Code stemmed not from a drop in model performance, but from bugs in the product layer and changes to prompt configuration.

The issue wasn't a single cause but multiple overlapping changes that compounded the problem.

• Reduced default reasoning effort: a measure taken to reduce latency complaints
• Context handling issues: affected the maintenance of conversation and task context
• Text limits between tool calls: adjustments to instruction style weakened long-horizon reasoning and multi-step edits

Anthropic has carried out fixes, rollbacks, and resets of subscriber usage limits for these issues.

The core message is clear. The quality of a coding agent is heavily influenced not just by model weights, but by orchestration, prompt policy, context management, and how tools are connected. So when performance degradation is observed, rather than first suspecting a model swap, one should look at the product layer's change history and release timeline together.

Prompted by this incident, the article also proposes the following operating principles.

• Track model versions and product configuration separately
• Establish explicit rollback criteria
• Use a canary channel to apply changes to major codebases first
• Manage prompt policy as a production configuration