Reliable and Flexible Task Planning Framework Via Agentic Collaboration
Abstract
Task planning requires balancing correctness with flexibility: classical planners guarantee soundness but demand formally specified domain models such as those based on the Planning Domain Definition Language (PDDL), while Large Language models (LLMs) interpret natural language yet produce unreliable plans. Existing LLMs-to-PDDL pipelines still rely on manually authored domain and example files, which are labor-intensive and error-prone. We present a Reliable and Flexible Task Planning Framework that achieves end-to-end planning from natural language alone through agentic collaboration. Our framework employs three specialized agents: a Domain Agent for PDDL domain generation with automatic validation and iterative correction, a Problem Agent with a three-stage context bootstrapping pipeline that eliminates manual example instances, and a Review Agent that diagnoses planning failures and routes corrective feedback to the responsible agent. Evaluated on seven IPC domains, our method achieves an 87.9% average success rate, outperforming LLM+P by 18.6%, with ablation studies confirming each component’s necessity.