An update on recent Claude Code quality reports

2026-04-24 · Source: Simon Willison's Weblog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Robotics & Autonomous Systems · Depth: Intermediate, quick

Summary

Anthropic has confirmed that a high volume of user complaints regarding degraded code quality from Claude Code over the past two months were valid. The issue stemmed not from the underlying models, but from three distinct problems within the Claude Code harness. One significant bug, introduced on March 26, involved a change intended to clear older session thinking after an hour of idleness to reduce latency. However, a defect caused this clearing to occur repeatedly during subsequent turns in the session, leading Claude to appear forgetful and repetitive to users. This particularly impacted users who frequently return to long-idle sessions, potentially affecting a substantial portion of user interactions.

Key takeaway

For AI Architects designing agentic systems, you should prioritize rigorous testing of your system's harness and session management logic. Complex interactions, like those involving session state and idle timeouts, can introduce subtle yet impactful bugs that mimic model performance issues. Ensure your testing protocols include scenarios for long-running, intermittently used sessions to prevent similar degradations in user experience.

Key insights

Harness-level bugs, not model issues, caused Claude Code's recent performance degradation.

Principles

• Harness bugs can mimic model failures.
• Session state management is critical.

In practice

• Implement robust session state logging.
• Test long-idle session resumption.

Topics

• Claude Code
• Anthropic
• AI Model Quality
• Agentic Systems
• Session Management

Best for: AI Architect, AI Engineer, MLOps Engineer, Prompt Engineer

Related on AIssential

• An update on recent Claude Code quality reports — Three distinct changes caused Claude Code's perceived degradation, highlighting the complexity of LLM product management.
• The TechBeat: Claude Code Works Better When You Let Sessions Die (6/6/2026) — Optimizing Claude code performance requires proper session termination.
• Applying an Agentic Coding Tool for Improving Published Algorithm Implementations — Agentic coding tools can significantly improve published algorithm implementations across diverse domains within a single workday.
• How CodeRabbit used Claude to build an agent orchestration system — Planning quality directly dictates AI code output quality, making early validation crucial to avoid expensive rework.
• Mystery solved: Anthropic reveals changes to Claude's harnesses and operating instructions likely caused degradation — Product-layer changes, not model degradation, caused Claude's perceived performance issues.
• v2.1.62 — Claude Code continuously evolves, focusing on performance, stability, and developer experience through new features and bug fixes.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Simon Willison's Weblog.