Online Monitoring and Corrective Steering of Programming Agents
Source
arxiv.org
Author
Shuyang Liu, Saman Dehghan, Ji Young Kim, Jatin Ganhotra, Martin Hirzel, Reyhaneh Jabbarvand
Date
Context
LivePlan helps a coding AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → recover from an unproductive repair attempt by separating problem detection from advice. A rule-based monitor watches the agent's actions for signs of drift or repetition. Only when it detects trouble does it call a LLMA large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.Full definition → to recommend a next step, concentrating additional model calls on runs that need intervention. In experiments using SWE-agent on two software-repair benchmarks, the authors report higher resolution rates, with additional successes concentrated on medium and hard problems. The contribution is a way to steer an ongoing repair without asking another model to reassess every step. Research abstract
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
articleCoupling Planning with Episodic Memory in LLM Agents for Software Issue ResolutionJiahao Zhang, Yifan Zhang, Yu Huang
postHow to really stop your agents from making the same mistakesGarry Tan
articleCodeGrep: An RL-Trained Retrieval Agent for LLM Coding AgentsWuya Chen, Yihao yang, Yang Cao, Yue Lin