AutoCompact: long-horizon coding agents learn when to compact—and what to keep
Oct 1 Zhang et al.: SFT+RL teaches when to compact, what working state to keep, and how to continue; +9.2 / +5.0 abs pass on SWE-bench Verified / SWE-PolyBench; 256K never overflows, 16K shares length-triggered fallback.
Source: AutoCompact: Learning When to Compact Context in Long-Horizon Coding Agents
Xuan Zhang, Longtao Zheng, Cunxiao Du, Bo An, and Xin Dong treat context management for repository agents as three coupled decisions: when to compact, what working state to preserve, and how to continue from that state—not merely how to avoid overflow. Length-triggered methods wait until a budget threshold; proactive rubrics and offline-inserted compression calls leave summaries and post-compaction actions unsupervised. AutoCompact adds a model-invocable compact() action, collects 1,052 judge-corrected on-policy trajectories (trigger / working-state / continuation corrections executed before the environment continues), fine-tunes, then runs outcome-based RL (GRPO, binary task success only) so coding and compaction share one reward.
On the same Qwen3-Coder-30B-A3B scaffold, AutoCompact reaches 39.6% on SWE-bench Verified and 24.5% on SWE-PolyBench Verified—+9.2 and +5.0 absolute over the full-history base (30.4 / 19.5). Gains hold across inference budgets from $0.10–$4.00 per task under a 256K window that never hits forced compaction and under a shared 16K length-triggered fallback. Ignoring the same trained checkpoint’s compaction calls drops pass rates (largest gap 19.9% at $0.10), so the win is not training alone. RL also raises proactive compact use (44.3% → 58.5% of tasks) while cutting key-state and next-action summary omissions.
Why it matters for code review
Continuous overnight lanes fill context with stale searches and failed hypotheses long before the window overflows; agents that only compact at the limit restart mid-stage or lose the next edit plan. Review agent PRs for AutoCompact-shaped contracts: a learned compact point after a stage resolves, a working-state summary that names files and remaining actions, and a first post-compact step that follows that plan rather than re-exploring. Always-on factories that FOMO for another pass need this to sustain verified momentum without context blowups—pair with Osmani on state, sessions, and handoffs and Li’s grow-the-harness-not-the-context when the merge gate asks whether the lane can run past a single context fill.