AgentDisCo

Towards Disentanglement and Collaboration in Open-ended Deep Research Agents

Jiarui Jin, Zexuan Yan, Shijian Wang, Wenxiang Jiao, Yuan Lu
Xiaohongshu Inc.

Project Page | arXiv | Code | GALA Benchmark | Renderer Gallery

---

Overview

AgentDisCo is a Disentangled and Collaborative architecture for open-ended deep research agents. Instead of coupling information exploration (search-query planning) and information exploitation (outline/report synthesis) inside one monolithic module, AgentDisCo separates them into a critic-generator research loop:

• The Critic Agent evaluates the evolving outline, identifies missing evidence, and emits gap-aware blueprints with targeted search queries.
• The Generator Agent retrieves new evidence, updates the outline, and maintains grounded references through a persistent document bank.
• The converged outline is passed to a Writer Agent for long-form synthesis and then to a Render Agent for HTML pages, posters, slides, and Rednote-style deliverables.

Highlights

• Disentangled exploration and exploitation. AgentDisCo formulates deep research as an adversarial-yet-collaborative optimization process between search-query generation and outline synthesis.
• Blueprint-centered iterative research. Each blueprint binds a report section to dedicated search queries, making retrieval more structured and incremental across rounds.
• Document bank for citation fidelity. Retrieved references are filtered, summarized, indexed, and carried across turns to reduce context noise and citation drift.
• Meta-optimization harness. Code-generation agents such as Claude Code or Codex can explore agent configurations and construct a reusable policy bank for search-query strategies.
• GALA benchmark. We introduce General AI Life Assistants (GALA), a lifestyle-oriented deep research benchmark mined from real user interaction histories.
• Multimodal render agent. Structured reports can be transformed into webpages, poster cards, slides, and Rednote-style content for end-user consumption.

Architecture

AgentDisCo runs a planner-to-render pipeline:

1. Planner Agent classifies the user query into intent categories and response style.
2. Critic Agent scores the current outline and proposes blueprints with targeted subqueries.
3. Generator Agent retrieves evidence and revises the outline plus references.
4. Document Bank keeps a citation-ready memory of useful evidence across optimization rounds.
5. Writer Agent expands the converged outline into a grounded Markdown report.
6. Render Agent packages the report into visual outputs such as HTML pages, posters, or slide decks.

Meta-Optimization Harness

AgentDisCo can optimize its own search strategy through an outer harness. In this setting, the generator agent is repurposed as a scoring agent that evaluates critic outputs and produces quality signals over search results. A code-generation agent then uses these signals to evolve reusable strategies in a policy bank, enabling systematic improvement of query generation across tasks and retrieval domains.

GALA Benchmark

Existing deep research benchmarks are often concentrated in academic, technical, or consulting-style queries. GALA targets a different regime: everyday information needs from real user behavior.

We mine latent deep research interests from Rednote user interaction histories, synthesize candidate queries with an agentic workflow, and distill a high-quality benchmark through LLM screening plus human verification. GALA emphasizes lifestyle categories such as Home & Hobbies, Travel, and Fashion & Beauty, complementing prior benchmarks like DeepResearchBench, DeepConsult, and DeepResearchGym.

Results

Using Gemini-2.5-Pro as the base model, AgentDisCo achieves competitive or leading performance across public deep research benchmarks:

• DeepResearchBench: AgentDisCo w/ Harness reaches 51.90 RACE overall, outperforming strong open and closed deep research baselines in the reported setting.
• DeepConsult: AgentDisCo w/ Harness achieves 66.86% win rate and 6.96 average score, surpassing compared systems under the pairwise evaluation protocol.
• DeepResearchGym: AgentDisCo w/ Harness reaches 96.77 overall, with strong performance on balance, breadth, support, depth, and insightfulness.
• GALA: AgentDisCo provides the reference reports and evaluation harness for lifestyle-oriented open-ended research queries.

Render Agent

The render agent converts structured research reports into multiple presentation modalities. It extracts salient points from Markdown or PDF reports, organizes them into template-specific assets, and renders the final result as webpages, posters, slide decks, or Rednote-style vertical images.

  

End-to-End Showcase

The system supports the full path from a user query to a refined outline, grounded report, and visual presentation. A running example follows a query about Japan's aging population from planning through critic-generator optimization, writing, and rendering.

Resources

• Project page: https://agentdisco-project.github.io/
• Paper PDF: https://agentdisco-project.github.io/static/pdfs/Deep_Research_XHS.pdf
• RACE data: https://agentdisco-project.github.io/race_data.jsonl
• Renderer gallery: https://agentdisco-project.github.io/#renderer-gallery

Citation

@misc{agentdisco,
      title={AgentDisCo: Towards Disentanglement and Collaboration in Open-ended Deep Research Agents}, 
      author={Jiarui Jin and Zexuan Yan and Shijian Wang and Wenxiang Jiao and Yuan Lu},
      year={2026},
      eprint={2605.11732},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2605.11732}, 
}