Anthropic explains Claude Code drop

Anthropic said on April 23 that three separate product changes, not a model downgrade, caused the recent drop in Claude Code quality. (anthropic.com) The company traced the regressions to Claude Code, the Claude Agent SDK, and Claude Cowork after a month of user reports about weaker coding help and broken workflows. (anthropic.com) (aol.com) The first change landed March 4, when Anthropic lowered Claude Code’s default reasoning effort from high to medium to reduce user-interface latency. The second shipped March 26, when a bug repeatedly cleared the model’s “thinking” history after tool calls. (anthropic.com) A third change on April 16 added a system prompt that pushed Claude to be more concise, and Anthropic said that made coding responses worse for some users. The company said it reverted or fixed the issues by April 20. (anthropic.com) The episode turned a fuzzy complaint into a concrete engineering problem: a coding assistant can look like the same product while hidden defaults, state handling, and prompts change underneath it. Anthropic said the core application programming interface and inference layer were not affected. (anthropic.com) Anthropic also reset usage limits for subscribers on April 23 and said it is changing how it rolls out defaults, tests stateful systems, and monitors quality regressions in coding products. (anthropic.com) At the same time, Anthropic put built-in memory for Claude Managed Agents into public beta. The feature lets agents keep information across sessions by reading and writing memory as files inside the agent filesystem. (claude.com) Anthropic said organizations can scope that memory to a workspace, inspect changes with audit trails and version history, and export memory files through an application programming interface. The company pitched that export option as a way to move data without being trapped inside one vendor’s system. (claude.com) Anthropic cited Rakuten as an early user, saying its long-running agents cut first-pass errors by 97%, lowered cost by 27%, and reduced latency by 34% with memory enabled. It also said Wisedocs sped document verification by 30%. (claude.com) Anthropic launched Claude Managed Agents in public beta on April 8 as a hosted way for developers to run autonomous agents without managing the infrastructure themselves. The new memory feature addresses one of the main limits of those systems: every session used to start from zero. (the-decoder.com) (claude.com) Taken together, the two announcements show Anthropic fixing a month-long reliability problem in one developer product while adding persistence to another. Both changes center on the same question: what an agent remembers, and what happens when that memory breaks. (anthropic.com) (claude.com)