Atlas · models
Compaction
Shrinking older context into a summary to make room. Details can be lost.
See it working: Plate I
In plain words
When the window fills up, something has to go. Compaction replaces older steps with a shorter summary.
How it works
Harnesses and APIs compact in different ways. Some write a readable summary; OpenAI’s Responses API can carry prior state as an opaque encrypted item.
Going deeper
Plate I’s one-line summary is a simplification chosen to show the cost: after compaction, the agent no longer has the exact text it read earlier. Compaction output is not a view into hidden reasoning.
Sources
- Compaction OpenAI
It is opaque and not intended to be human-interpretable.
Quotes were checked against the live pages on 2026-11-08.