An update on recent Claude Code quality reports
Simon Willison 行业观点 进阶 Impact: 7/10
The culprit behind Claude Code's quality decline over the past two months wasn't model degradation, but three harness-level bugs, with a 'session state cleanup' glitch exposing hidden complexities in AI Agent engineering.
Key Points
- Anthropic confirmed: Recent Claude Code quality issues stemmed from three harness bugs
- not model capability degradation
- The critical bug was session state mismanagement: intended to clear old thinking after 1-hour idle to reduce latency
- but erroneously repeated every turn
- causing AI 'amnesia
- This reveals core challenges in Agent engineering: scaffolding systems around LLMs (context management
- state maintenance
- tool orchestration) are more prone to subtle systemic failures than the models themselves
- Long-lived sessions are high-frequency scenarios for Agent products
- but engineering implementations often assume short-term interactions—this mismatch is a hidden killer of Agent reliability
- Debugging Agent systems requires distinguishing 'model uncertainty' from 'engineering deterministic bugs
- with the latter often masquerading as the former
- leading to misattribution
Analysis
"For the past two months, the Claude Code user community has been filled with uneasy complaints: the former 'coding god' seemed to have gotten dumber, repeating itself, forgetting previous context, and giving noticeably worse advice. Many suspected Anthropic had secretly swapped model versions, or that Claude 3.7 Sonnet itself had degraded.
Analysis generated by BitByAI · Read original English article