AutoCompact trains agents to compact context themselves, +9.2 on SWE-bench Verified
omarsar0 · x · 2026-10-04
AutoCompact trains agents to natively decide when to compact context, what working state to keep, and how to resume. A judge reviews and corrects the base agent's compaction decisions, corrected trajectories feed SFT, then RL with task-success rewards trains coding and compaction jointly. Pass rates improve by 9.2 points on SWE-bench Verified and 5.0 on SWE-PolyBench Verified — gains that hold even with a never-overflowing 256K window. Part of a rising trend (AutoHarness, AutoContext, Meta's context-management paper) of models natively absorbing harness duties.
Related event: AutoCompact Lets Agents Decide Context Compression Themselves(2 posts)→
More from coding & agent
- Devs flag GPT's defensive programming for hiding bugs; OpenAI asks for /feedback cases — pvncher · 2026-11-19
- Flutter ships A2UI 1.0: agents respond with real interactive widgets, not chat text — rseroter · 2026-11-19
- Skills vs MCP vs RAG vs Memory: the four context mechanisms AI agents actually need — MaryamMiradi · 2026-11-19
- Anthropic details agent memory: 'dreaming' lets Claude organize its experience — daniel_mac8 · 2026-11-19
- The case for externalized memory as the real fix for continual learning — daniel_mac8 · 2026-11-19
- FT: US tech outages blamed on rush to fix bugs found by AI — nordicinst · 2026-11-19