Anthropic published this postmortem after users experienced a noticeable drop in Claude Code quality.
That’s the important setup: the issue was real, but the root cause mostly wasn’t the base model.
Three levels shape the outcome:
- Model — the base intelligence
- Context — what the agent remembers, what gets dropped, and how state carries across turns
- Harness — the prompts, tool wiring, defaults, and guardrails around the model
In this case, the quality drop mostly came from layers 2 and 3:
less reasoning by default, lost context across turns, and prompt changes that reduced coding quality.
Same model family.
Different orchestration.
Very different user experience.
That’s why agent quality won’t be defined only by which model you use.
It will be defined by how well we engineer the system around it.
Postmortem link in the first comment.