Auto-compaction
person concept tool org talk claim — click a node to jump to its page; hover an arrow for the relation
Auto-compaction is a context-management technique used in AI agent engineering that automatically compresses or summarizes conversation/context history to prevent hitting context window limits during long-running agent sessions.
Role in Harness Engineering
Ryan Lopopolo (OpenAI) identifies auto-compaction as a component of Harness Engineering — the broader practice of building software where humans steer and agents execute. He notes it as a capability that has been continuously improving, citing GPT-4.5 and CEX (presumably a context-extension system) as performing particularly well at it. 11:20
The framing as a "continued improvement" suggests auto-compaction is treated as an ongoing engineering challenge rather than a solved problem — one where model and system quality directly affects how effectively long-horizon agent tasks can be sustained without losing critical context.
Significance
For agentic systems that must maintain coherent state across many tool calls, code edits, or multi-step reasoning chains, auto-compaction is a practical prerequisite for extended autonomous operation. Without it, agents would hit context limits and require manual intervention, undermining the "agents execute" half of the human-steers/agent-executes model.