TL;DR
AgentDisCo is a novel deep research agent architecture that decouples information exploration (searching) from information exploitation (writing) into an adversarial-yet-collaborative loop between a critic and a generator agent. Driven by an automated meta-optimization harness, it outperforms leading closed-source systems on standard benchmarks and introduces GALA, a new benchmark tailored to everyday lifestyle research needs.
What Problem Does This Paper Address?
Building autonomous AI agents capable of executing "deep research"—gathering web-scale information and synthesizing it into comprehensive reports—is a major frontier in LLM development. However, existing architectures suffer from a persistent bottleneck: they entangle information exploration (planning and formulating search queries) with information exploitation (generating structured outlines or reports) within a single, undifferentiated module.
This entanglement leads to two major issues:
- Lack of informational incrementality: The agent struggles to systematically build upon what it has already discovered.
- Unstable refinement: Without explicit optimization objectives, the agent frequently oscillates between over-revising and under-revising its outlines without ever converging on a high-quality structure.
Furthermore, existing deep research benchmarks focus heavily on academic or domain-specific consulting queries, failing to capture the organic, lifestyle-oriented information needs of everyday users.
Methods
AgentDisCo addresses these flaws by formulating deep research as a dual-agent cooperative Markov Decision Process (MDP).
The architecture splits the core workload into two specialized, interacting roles:
- The Critic Agent: Evaluates the completeness and quality of the current outline (the exploitation state) and generates targeted, gap-aware search queries wrapped in structured "report blueprints."
- The Generator Agent: Takes the critic's queries, retrieves updated search results, and refines the outline and references (the exploration state).
To prevent context window bloat and manage references across multiple rounds, the authors introduce a Document Bank—a lightweight tracker that parses documents into fine-grained snippets, scores them for relevance, and filters out noise before they reach the generator.
Meta-Optimization Harness
To automatically refine how the agent searches, the authors implement an outer meta-optimization loop. A code-generation agent (like Claude-Code) systematically explores different agent configurations. During this process, the generator agent is repurposed as a Score Agent to evaluate the critic's query quality across dimensions like completeness, diversity, and search coverage. This allows AgentDisCo to autonomously build a Policy Bank of reusable search strategies without human intervention.
Datasets
The authors evaluate AgentDisCo on three established benchmarks:
- DeepResearchBench: 100 PhD-level complex research tasks.
- DeepConsult: A collection of business and consulting prompts.
- DeepResearchGym: 100 queries sampled from the large-scale Researchy Questions dataset.
The GALA Benchmark
To bridge the gap between academic tasks and real-world utility, the authors introduce GALA (General AI Life Assistants). Mined from the historical browsing and commenting behaviors of 10,000 active users on the Rednote (Xiaohongshu) platform, GALA consists of 100 curated, highly specific lifestyle queries spanning domains like Home & Hobbies, Travel, and Fashion & Beauty.
Key Insights and Ideas
- Outlines Over Reports: Iterative optimization is far more effective when performed on intermediate outlines rather than full reports. Outlines offer superior structural flexibility and make context management significantly easier.
- The Minimax Synergy: While the critic and generator act adversarially—the critic probing for missing evidence and the generator defensively consolidating the outline—they share a global cooperative reward. This push-and-pull naturally drives the system toward comprehensive coverage.
- Query Generation is the True Bottleneck: LLMs excel at summarizing and extracting information but are comparatively weak at formulating effective search queries across diverse platforms. Focusing meta-optimization strictly on query generation yields the highest performance gains.
- Community Data Superiority: For lifestyle-oriented queries, retrieving data from community-grounded platforms (like Rednote) yields more practical, experience-rich, and user-aligned reports than generic web searches.
Key Results
AgentDisCo achieves state-of-the-art performance across multiple deep research benchmarks using Gemini-2.5-Pro as its base model:
- DeepResearchBench: AgentDisCo with the meta-optimization harness achieved an overall RACE score of 52.11, outperforming Gemini-2.5-Pro-Deepresearch (49.71) and closed-source alternatives like OpenAI-DeepResearch (46.45). When scaled to Claude-Opus-4.6, it reached 54.02.
- DeepConsult & DeepResearchGym: It achieved an overall score of 6.86 on DeepConsult and 96.21 on DeepResearchGym, demonstrating robust generalization.
- GALA Benchmark: AgentDisCo consistently beat proprietary baselines like Doubao-Research and OpenAI o3-DeepResearch. Notably, AgentDisCo performed better when utilizing the Rednote search engine alone (51.02) than when using Google Search alone (50.00) for these lifestyle tasks, proving the value of domain-specific content.
Limitations and Future Work
While AgentDisCo represents a major step forward in agent collaboration, the authors note that their automated harness optimization currently focuses exclusively on search query generation. The downstream tasks—specifically outline organization and the final report writing phase handled by the Writer Agent—remain handcrafted. The authors leave the systematic meta-optimization of these writing and structuring modules to future work.
Summary updated