TLDR: ProSEA is a novel multi-agent AI framework designed for iterative problem solving. It features a hierarchical architecture with a Manager Agent orchestrating specialized Expert Agents. Unlike other systems, ProSEA’s Expert Agents provide detailed feedback on failures and new constraints, enabling the Manager to adaptively refine plans. This allows for ‘two-dimensional exploration’ (breadth by Manager, depth by Experts) and makes the system model-agnostic and capable of seamless human collaboration. Experiments on FinanceBench show ProSEA outperforms state-of-the-art baselines autonomously, demonstrating its potential for more adaptive and human-aligned AI.
Artificial intelligence agents have made significant strides in tackling complex tasks, thanks to advancements in large language models (LLMs). However, many existing AI systems still grapple with limitations such as static planning, rigid interactions, and a lack of adaptive reasoning. They often struggle with complex, multifaceted problems that demand deep reasoning and flexible problem-solving strategies, frequently failing to learn effectively from past mistakes.
A new research paper introduces a groundbreaking framework called ProSEA, which stands for Problem Solving via Exploration Agents. Developed by William Nguyen, Vinh Luong, and Christopher Nguyen from Aitomatic, Inc., ProSEA aims to overcome these limitations by offering a modular, general-purpose multi-agent system designed for iterative problem solving through continuous exploration and dynamic plan evolution. You can read the full paper here.
A Hierarchical Approach to Problem Solving
ProSEA features a hierarchical architecture that mirrors effective organizational structures. At its core is a Manager Agent, which acts as the central orchestrator. This Manager Agent is responsible for overseeing and coordinating multiple specialized Expert Agents. It breaks down complex tasks into smaller, manageable subtasks and assigns them to the appropriate Expert Agents.
Unlike previous systems, ProSEA’s Expert Agents provide rich, structured feedback. When an attempt fails, they don’t just report failure; they detail the reasons why, what new constraints were discovered, and what alternative approaches were explored. This detailed feedback is crucial, enabling the Manager Agent to adaptively refine its plan based on real-time insights gained from exploratory traces.
Two-Dimensional Exploration and Adaptive Reasoning
One of ProSEA’s core innovations is its ability to perform two-dimensional exploration. The Manager Agent explores the solution space in breadth by decomposing tasks and dynamically replanning based on feedback. Simultaneously, the Expert Agents delve into depth, engaging in iterative reasoning within their specialized domains. This division of labor creates a robust and modular system where global strategy and local expertise complement each other, allowing ProSEA to tackle problems beyond the scope of single-agent systems.
The framework is also model-agnostic, meaning it can be powered by any suitable LLM without requiring additional training or fine-tuning. Furthermore, ProSEA is designed for seamless integration with human collaborators. It operates autonomously by default but can incorporate human feedback at any stage of the reasoning process, treating human input as another valuable source of expert knowledge.
How ProSEA Works: The Pipeline
The ProSEA pipeline begins when the Manager Agent receives an initial question. This question is forwarded to a Problem Analyzer for preliminary analysis, which extracts key constraints and requirements. The analyzed problem then goes to a Planner module, which generates a multi-step solution plan. Each step is assigned to a domain-specific Expert Agent for execution. These Expert Agents use their specialized knowledge, external tools, and internal reasoning loops to achieve their assigned goals. The results and detailed feedback from each step are sent back to the Manager Agent, which evaluates the need for plan adjustments. This iterative process continues until a complete solution is found, or the problem space is sufficiently explored.
Impressive Performance on FinanceBench
To validate its effectiveness, ProSEA was rigorously tested on the challenging FinanceBench dataset, which includes 150 complex financial questions. In fully autonomous mode, without any human collaboration, ProSEA demonstrated state-of-the-art performance. It significantly outperformed traditional approaches like LlamaIndex RAG agents, LangChain ReAct agents, and OpenAI Assistants across various difficulty levels.
Notably, ProSEA achieved new benchmarks in categories such as information retrieval, comparative analysis, change calculations, and factor explanation. While it performed comparably to DANA (another advanced system) in complex calculations, ProSEA’s key advantage is its autonomy. DANA requires extensive domain knowledge engineering and manual plan guidance from human experts, whereas ProSEA operates entirely on its own, highlighting its superior scalability and practical applicability.
Also Read:
- FlowSearch: A Multi-Agent System for Adaptive Deep Research
- Double-Loop Multi-Agent Framework Automates Scientific Research with Evolving Plans and Dynamic Execution
The Future of Adaptive AI Agents
ProSEA represents a significant leap forward in multi-agent systems for complex problem-solving. By combining hierarchical task decomposition with rich, structured feedback mechanisms and two-dimensional exploration, it offers a powerful framework that can adaptively refine its approaches based on discovered constraints and failures. This allows the system to navigate obstacles and find viable solutions without relying on predetermined paths or constant human intervention.
The strong autonomous performance of ProSEA suggests its potential as a foundation for more transparent, adaptive, and human-aligned AI agents, capable of serving as true collaborative partners in real-world scenarios.
