In April 2026, Anthropic published a postmortem explaining a noticeable drop in quality in Claude Code over several weeks. Users reported worse responses, loss of context, and inconsistent behavior — and they were right.
What actually happened
The degradation was not caused by changes to the model itself.
Instead, three separate changes in the agent "harness" (the system around the model) combined in unexpected ways:
Reduced reasoning effort
- Default reasoning was lowered from high to medium to improve latency.
- This made responses faster, but less thoughtful.
- Later reverted after user feedback.
Session memory bug
- A fix intended to clear old context after inactivity misfired.
- Instead of running once, it triggered every turn.
- Result: the model appeared forgetful and repetitive.
Additional system-level changes
Smaller adjustments (e.g., prompt / behavior tuning) were deployed.
Individually minor, but together amplified the degradation.
👉 None of these changes were critical alone.
👉 Together, they created a system-wide quality drop.
Key insight
The model didn’t get worse — the surrounding system did.
This is an important shift:
Agent quality depends as much on infrastructure as on the model itself.
Why it was hard to detect
Changes were rolled out at different times
Affected different parts of the system (Code, SDK, Cowork)
Symptoms were inconsistent across users
This made the issue look like “random degradation” instead of a clear regression.
What was fixed
All three issues were identified and reverted or fixed by April 20. :contentReference[oaicite:1]{index=1}
Reasoning effort restored
Memory bug fixed
System behavior stabilized
Lessons learned
1. Small changes can compound
Even minor optimizations (latency, memory, prompts) can interact in unexpected ways.
2. Latency vs quality is a real tradeoff
Optimizing for speed can silently reduce reasoning depth.
3. Agent systems are fragile
The “wrapper” around the model (memory, prompts, orchestration) is just as critical as the model itself.
4. Observability matters
Without clear metrics, degradations can go unnoticed or be misdiagnosed.
Takeaway
If you’re building AI agents:
- Don’t assume the model is the problem
- Monitor system-level changes carefully
- Treat agent infrastructure as production-critical
- The intelligence is not just in the model — it’s in the system around it.
/ revisions
- Published · 29 Apr 2026
- POST-001
- CC BY 4.0