Member-only story
The same agent that refused a prohibited action nine times in a row did it on the tenth — not because the model changed, but because the harness summarized the rule out of its context. With the policy fully visible, violation rate was 0%. After one compaction pass, it hit 59% on DeepSeek-V4-Flash and 41% on GPT-5.4-mini. Run the same setup through AutoGen’s stock BufferedChatCompletionContext
and it hits 100%.
That’s from ConstraintRot, a 1,323-episode benchmark in a paper called Governance Decay (arXiv:2606.22528). I spent yesterday pulling it apart, cross-checking it against the compaction code that ships in LangGraph, LangMem, the OpenAI Agents SDK, and Anthropic’s own context-editing API — and then rebuilding the mitigation as ~40 lines you can drop into an existing agent today.
The uncomfortable part isn’t that summarization is lossy. Everyone knows summarization is lossy. It’s what it preferentially loses, and the fact that the loss is completely silent.
Why this is showing up now #
Every serious agent harness in production compacts. It has to — a long-horizon coding or ops session blows past the window in an afternoon.
Claude Code auto-compacts and has a manual/compact
. The most-upvoted issues in…