LivePlan boosts AI coding agents' fix rates by 15.2% for $0.08
New AI monitor catches drifting agents and steers them back on track
Fixing GitHub issues in large-scale codebases is a long-horizon task where AI agents often drift off course, repeat failed actions, or give up without a working patch. Researchers Shuyang Liu and colleagues from IBM Research and UIUC propose LivePlan, a system that monitors and corrects programming agents in real time. The key insight is decoupling judging from advising: a deterministic, rule-based monitor scans agent trajectories for signs of inefficiency or drift without invoking an LLM. Only when an anomaly is detected does LivePlan consult an advisor LLM for a high-level, next-step correction. This avoids the costly and misleading re-planning of prior approaches.
Built on top of SWE-agent, LivePlan was evaluated across SWE-bench Verified and SWE-bench Pro using five LLMs (three executors, two advisors). Compared to vanilla SWE-agent, it achieved consistent resolution-rate gains of up to 15.2% (average 9.9%) while adding only $0.08 per instance. The improvements concentrated on medium and hard problems, and LivePlan even solved issues that no baseline could fix, with minimal regression on already-successful runs. The paper is available on arXiv (2608.06701).
- LivePlan decouples a deterministic rule-based monitor from an LLM advisor, cutting intervention costs
- Improved SWE-bench issue resolution by 9.9% average (15.2% max) over vanilla SWE-agent
- Each corrected instance costs just $0.08, with new solutions on hard problems no baseline solved
Why It Matters
Cheap, real-time agent monitoring could make AI coding assistants far more reliable for production debugging and multi-file fixes.