Autocompact
Compaction triggered automatically by the harness when the context window approaches full.
Autocompact is compaction the harness triggers on its own, when the context window approaches full, without the user asking. It summarises the session so far, drops the raw transcript, and continues with the summary plus the recent turns.
It exists because the alternative is worse. A context window that silently overflows fails the request outright, and a harness that leaves the choice to the user trains people to ignore the warning until the failure happens.
The cost of doing it automatically is the loss of detail. A summary cannot carry every fact from a hundred tool calls, and what gets dropped is rarely the least important thing. The mitigation is a good summary format: decisions made and why, facts discovered, files touched, and what is still open. A prose summary of a long session keeps the narrative and loses the specifics, and the specifics are usually what you need.
Related terms
Compaction
Summarising the old conversation so the new part fits in the window.
Context Window
Everything the model sees on one request. Finite, and it decides what fits.
Context Degradation
The slow loss of quality as a session grows, even though nothing broke.
Handoff
Moving work from one session, or one agent, to another without losing the thread.
Memory
A system's ability to carry information from one request to the next.
Want this working inside your team?
It starts with one conversation about your workflow.