Enterprise AI Analysis

The Prog-Act (Progressive reasoning and Acting) approach is a semi-dynamic planning method that enhances multi-agent collaboration by enabling active assistance-seeking, fixed-point verification, and escalating collaboration through a dual-loop error detection and correction mechanism.

This dual-loop mechanism (localized fixes and global revisions) optimizes error handling while preserving task continuity, balancing adaptability and efficiency.

A Graph Memory structure further enhances error traceability and cross-step coordination by modeling tool/subtask dependencies.

Effective coordination of specialized tools is crucial for complex multi-step tasks. Existing methods often rely on explicit parameter passing, neglecting implicit dependencies like runtime artifacts (function objects, class instances).

PDE-Agent introduces a Resource-Pool, a dynamic repository for shared resources, which centralizes implicit parameter management, resolving inter-tool dependency gaps and enabling seamless multi-tool coordination.

To validate PDE-Agent, a new benchmark called PDE-Bench (or PDE-Data) was developed, featuring 100 differential equation test cases with full workflow annotations.

Multi-level evaluation metrics were proposed, encompassing global task completion (Pass@k), local tool invocation assessment (Semantic Textual Similarity, BERTScore), and logical collaboration process assessment (graph-based metrics like Subgraph Isomorphism, PageRank-JS, Graph Edit Distance, and Graph Embedding Similarity).

Experiments show PDE-Agent achieves a 90% overall success rate, significantly outperforming baselines like OctoTools (62%), OpenAlita (59%), and LLM-DS (56%).