Summary

AgentDisCo addresses the challenge of open-ended deep research by disentangling information exploration and exploitation within AI agents. It targets researchers and developers building sophisticated AI systems for complex information synthesis, offering a structured, iterative, and grounded approach that culminates in multimodal outputs. The architecture aims to improve the fidelity and efficiency of deep research processes.

How It Works

AgentDisCo employs a critic-generator research loop, treating deep research as an adversarial yet collaborative optimization process. A Critic Agent evaluates evolving outlines, identifies information gaps, and generates targeted search queries within "blueprints." A Generator Agent then retrieves evidence, updates the outline, and maintains grounded references in a persistent document bank, ensuring citation fidelity across iterative rounds. This disentangled approach allows for more structured retrieval and synthesis, while a meta-optimization harness can further refine search strategies using code-generation agents.

Quick Start & Requirements

Key resources include the Project Page, arXiv paper, and Renderer Gallery. The "Code" and "GALA Benchmark" links are mentioned but not explicitly provided with URLs in the text; the project's GitHub repository is assumed for code access. The system relies on large language models, with Gemini-2.5-Pro cited as a base. Specific installation commands, hardware prerequisites (e.g., GPU, CUDA), or setup time estimates are not detailed.

Highlighted Details

• Achieves competitive or leading performance on established deep research benchmarks: 51.90 RACE on DeepResearchBench, 66.86% win rate on DeepConsult, and 96.77 overall on DeepResearchGym.
• Introduces the GALA benchmark, focusing on lifestyle-oriented deep research queries derived from real user interactions.
• Features a multimodal Render Agent capable of transforming structured reports into webpages, posters, slide decks, and Rednote-style content.
• Incorporates a Meta-Optimization Harness for systematically improving search query generation strategies.

Maintenance & Community

The provided text does not detail community channels (e.g., Discord, Slack), specific maintenance plans, or a public roadmap. Author information is available via the arXiv paper.

Licensing & Compatibility

No specific open-source license is mentioned in the provided README content. Compatibility for commercial use or integration with closed-source systems is therefore not specified.

Limitations & Caveats

The provided description focuses on the system's capabilities and performance, and does not explicitly state any limitations, known bugs, or areas of ongoing development.