AutoCompact trains coding agents to decide when to compact context

AutoCompact’s authors say their agent learns when to compact context, what to keep and how to resume. They use judge-corrected runs for supervised fine-tuning, then train coding and compaction together with reinforcement learning. They report pass-rate gains over the base model of 9.2 points on SWE-bench Verified and 5.0 on SWE-PolyBench Verified, even with a 256K window that never overflowed. A commentator says how well the approach scales and works across harnesses remains unclear.