Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ResearStudio: A Human-Intervenable Framework for Building Controllable Deep-Research Agents

Linyi Yang, Yixuan Weng

TL;DR

The paper tackles the rigidity of current Deep Research agents that operate in a fire-and-forget mode by introducing ResearStudio, an open-source framework that enacts a Collaborative Workshop with a Planner–Executor core and a live plan-as-document interface for real-time human intervention. It describes a three-layer architecture and a unified tool ecosystem (document processing, search, code, and multi-modal capabilities) managed through a bidirectional protocol that streams updates to a live workspace. Empirically, ResearStudio delivers state-of-the-art performance on the GAIA benchmark while preserving transparent human-in-the-loop control, enabling interventions such as pausing, plan editing, and direct code adjustments. The results suggest that strong autonomous performance and fine-grained human control can coexist, with implications for safer, more trustworthy Deep Research agents and a pathway for broader, open-source development. The framework is released with full code, protocol, and demos to spur further research and safe deployment of controllable AI researchers.

Abstract

Current deep-research agents run in a ''fire-and-forget'' mode: once started, they give users no way to fix errors or add expert knowledge during execution. We present ResearStudio, the first open-source framework that places real-time human control at its core. The system follows a Collaborative Workshop design. A hierarchical Planner-Executor writes every step to a live ''plan-as-document,'' a fast communication layer streams each action, file change, and tool call to a web interface. At any moment, the user can pause the run, edit the plan or code, run custom commands, and resume -- switching smoothly between AI-led, human-assisted and human-led, AI-assisted modes. In fully autonomous mode, ResearStudio achieves state-of-the-art results on the GAIA benchmark, surpassing systems like OpenAI's DeepResearch and Manus. These results show that strong automated performance and fine-grained human control can coexist. The full code, protocol, and evaluation scripts are available at https://github.com/ResearAI/ResearStudio. We will continue to update the repository to encourage further work on safe and controllable research agents. Our live demo is publicly accessible at http://ai-researcher.net:3000/. We support the development of DeepScientist, which can be accessed at https://github.com/ResearAI/DeepScientist.

ResearStudio: A Human-Intervenable Framework for Building Controllable Deep-Research Agents

TL;DR

The paper tackles the rigidity of current Deep Research agents that operate in a fire-and-forget mode by introducing ResearStudio, an open-source framework that enacts a Collaborative Workshop with a Planner–Executor core and a live plan-as-document interface for real-time human intervention. It describes a three-layer architecture and a unified tool ecosystem (document processing, search, code, and multi-modal capabilities) managed through a bidirectional protocol that streams updates to a live workspace. Empirically, ResearStudio delivers state-of-the-art performance on the GAIA benchmark while preserving transparent human-in-the-loop control, enabling interventions such as pausing, plan editing, and direct code adjustments. The results suggest that strong autonomous performance and fine-grained human control can coexist, with implications for safer, more trustworthy Deep Research agents and a pathway for broader, open-source development. The framework is released with full code, protocol, and demos to spur further research and safe deployment of controllable AI researchers.

Abstract

Current deep-research agents run in a ''fire-and-forget'' mode: once started, they give users no way to fix errors or add expert knowledge during execution. We present ResearStudio, the first open-source framework that places real-time human control at its core. The system follows a Collaborative Workshop design. A hierarchical Planner-Executor writes every step to a live ''plan-as-document,'' a fast communication layer streams each action, file change, and tool call to a web interface. At any moment, the user can pause the run, edit the plan or code, run custom commands, and resume -- switching smoothly between AI-led, human-assisted and human-led, AI-assisted modes. In fully autonomous mode, ResearStudio achieves state-of-the-art results on the GAIA benchmark, surpassing systems like OpenAI's DeepResearch and Manus. These results show that strong automated performance and fine-grained human control can coexist. The full code, protocol, and evaluation scripts are available at https://github.com/ResearAI/ResearStudio. We will continue to update the repository to encourage further work on safe and controllable research agents. Our live demo is publicly accessible at http://ai-researcher.net:3000/. We support the development of DeepScientist, which can be accessed at https://github.com/ResearAI/DeepScientist.
Paper Structure (19 sections, 1 equation, 5 figures, 3 tables)

This paper contains 19 sections, 1 equation, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. The ResearStudio Framework
  4. Tool Usage
  5. Document-Processing Toolkit.
  6. Search Toolkit.
  7. Deliberate Browser Omission.
  8. Code Toolkit.
  9. Interactive Web Interface.
  10. Bidirectional Protocols
  11. Backend Models
  12. Experiments
  13. Overall Results.
  14. Discussion
  15. Conclusion
  16. ...and 4 more sections

Figures (5)

  • Figure 1: The overall architecture of the ResearStudio framework. This diagram illustrates the three core layers (L-1, L-2, L-3) and the primary workflow. The L-2 Agent Core, composed of a Planner and an Executor, processes a user's request. The Executor carries out steps by using tools from the L-1 MCP Toolbox. The entire process is accessible to the user via the L-3 WebPage, linked by a central Communication Protocol.
  • Figure 2: Core interaction workflows of the ResearStudio communication protocol. This diagram details the sequence of messages between the User, Planner, Executor, System (Toolbox), and Workspace for key operations, illustrating the bidirectional flow of information that enables real-time collaboration.
  • Figure 3: An overview of the ResearStudio user interface, designed as an integrated workspace for human-agent partnership. The layout provides a comprehensive view by juxtaposing the agent's real-time execution trace in the activity log with the tangible project artifacts in the file explorer and editor. This design allows a user to seamlessly monitor autonomous operations while retaining the ability to directly interact with and manage all project files and system states through dedicated controls.
  • Figure 4: A successful execution trace of a GAIA Level-3 computational puzzle within the ResearStudio UI, showcasing the agent's ability to write and execute code to find a correct solution.
  • Figure 5: An execution trace of a failed GAIA Level-3 physics problem. The agent correctly structures the problem but fails by selecting the wrong temperature (400°C), leading to a cascade of incorrect calculations.