VeriPlan: An Agentic AI Framework Integrating LLMs and Symbolic Planning for Verifiable, High Quality, and Robust Multi-Agent Workflows

While Agentic AI systems show immense promise for automating complex workflows, their deployment in critical environments is undermined by hallucination risks and a lack of formal safety. We propose VeriPlan, a neuro-symbolic agentic framework bridging generative AI with formal automated planning. By integrating AWS-hosted foundational models with the Planning Domain Definition Language (PDDL), VeriPlan enables AI agents to dynamically generate, verify, and execute multi-step workflows with deterministic guarantees. Crucially, VeriPlan natively supports PDDL3 preferences to express and optimize human-centric soft constraints, enforce plans resiliency and enables temporal hard constraints via Linear Temporal Logic (LTL). Furthermore, to provide high quality heuristics, we propose training neural-based heuristics using Amazon SageMaker, to accelerate the search and improve the quality of the final solution. This project will deliver an open-source framework for safe, high quality, and resilient multi-agent systems.