ManuSearch: A Transparent and Open Multi-Agent Framework for Deep Search in LLMs

Forum topic · 小凯 · 2026-07-05

Summary

ManuSearch is an open-source, transparent multi-agent framework for deep search in large language models, presented in a May 2025 arXiv paper (arXiv:2505.18105) by Lisheng Huang, Yichen Liu, Jinhao Jiang, Rongxiang Zhang, Jiahao Yan, Junyi Li and colleagues. The work addresses the opacity of commercial deep research systems, which combine reasoning, online search, and code execution behind closed APIs, making it hard for researchers to study or build on them. ManuSearch decomposes deep search into three collaborating agents: a planning agent that orchestrates the reasoning process, a search agent that interacts with the open web, and a code agent that executes programs for computation and data analysis. The framework is fully open, allowing researchers to inspect, reproduce, and extend the entire agentic search pipeline. According to the authors, ManuSearch outperforms established deep research systems on challenging benchmarks, demonstrating that transparent community-driven designs can rival proprietary solutions. The paper also provides analysis of agent orchestration, tool use, and evaluation protocols, and positions ManuSearch as infrastructure for democratizing deep research capabilities. It is relevant to readers interested in agentic retrieval-augmented generation, multi-agent collaboration, and open reproduction of commercial-grade deep search systems.

ManuSearch: Democratizing Deep Search in Large Language Models with a Transparent and Open Multi-Agent Framework

Source: arXiv:2505.18105 · May 2025 Authors: Lisheng Huang, Yichen Liu, Jinhao Jiang, Rongxiang Zhang, Jiahao Yan, Junyi Li, et al. (7 authors)

Overview

ManuSearch is a transparent, fully open multi-agent framework for deep search with large language models. It was motivated by the observation that commercial deep research systems (which interleave reasoning, online search, and code execution behind closed APIs) are powerful but opaque, limiting scientific study and community reuse. ManuSearch opens up the complete agentic search pipeline so that it can be inspected, reproduced, and extended.

Design: Three Collaborative Agents

Planning agent — orchestrates the overall deep research process, deciding when to search, when to execute code, and how to synthesize intermediate results.
Search agent — interacts with the open web to retrieve up-to-date external knowledge beyond the model's training data.
Code agent — writes and executes programs for computation, structured data processing, and verification tasks.

This decomposition mirrors the capabilities that closed deep research products combine internally (reasoning, search, and code execution), but exposes each component openly.

Key Claims

ManuSearch reportedly outperforms existing deep research systems (including closed commercial ones) on challenging deep search benchmarks.
The framework demonstrates that transparent, open designs can match proprietary agentic search pipelines, supporting reproducible research on agentic retrieval.
By exposing the full trajectory of planning, searching, and coding, ManuSearch enables study of agent orchestration, tool-use policies, and evaluation in agentic RAG systems.

Context

The paper sits at the intersection of agentic search, retrieval-augmented generation (RAG), and multi-agent collaboration. It complements related surveys and systems on deep research and LLM-based search agents, such as:

A Comprehensive Survey of Deep Research (arXiv:2506.12594)
A Survey of LLM-based Deep Search Agents (arXiv:2508.05668)
Agentic Reasoning (arXiv:2502.04644)

Takeaways

For researchers: ManuSearch provides an open baseline for studying how planning, retrieval, and execution should be coordinated in deep research agents.
For engineers: the modular three-agent design is a reference architecture for building agentic search stacks with pluggable retrievers and tool interfaces.
For the community: open, reproducible alternatives to closed deep research APIs help close the gap between academic evaluation and industrial systems.

For exact benchmark numbers and implementation details, refer to the original paper.