AgentDisCo: Disentanglement and Collaboration for Open-ended Deep Research Agents
In one sentence: AgentDisCo (Disentanglement and Collaboration) reframes open-ended deep research as an adversarial optimization problem between information "exploration" and "exploitation", explicitly separating "what to search for" and "how to use what is found"—usually crammed into a single module—and handing them to two agents (Critic / Generator) that play against each other and even swap roles to critique one another in a collaborative loop; it then uses a code-generated meta-optimization shell to automatically search agent configurations and distill a reusable "policy bank", matching or surpassing leading closed-source systems on multiple deep research report benchmarks. Year proposed: 2026 (arXiv:2605.11732, v1 2026-05; v2 2026-06) · Authors: Jiarui Jin, Zexuan Yan, Shijian Wang, Wenxiang Jiao, Yuan Lu (affiliations per the original paper) · Evaluation backbone model: Gemini-2.5-Pro Prerequisite reading: Deep Research overview · Multi-agent · STORM / Co-STORM
The problem: exploration and exploitation tangled together
Most deep research agents stuff "discovering new information" and "organizing existing information into a report" into the same module, done off-handedly by the same loop. AgentDisCo argues this is a structural flaw:
- Exploration cares about "what is still missing, where should I search next"—the goal is to broaden information coverage and fill knowledge gaps.
- Exploitation cares about "how to trim and organize existing material into a well-argued, high-quality report"—the goal is to converge and maximize quality.
These two tasks have naturally conflicting optimization directions: one wants to diverge, the other to converge. When tangled together, an agent tends to either search a lot but organize poorly, or write smoothly but cover shallowly, and is hard to tune in a targeted way. AgentDisCo's core claim is: disentangle exploration and exploitation into independent components, then let them correct each other in an adversarial + collaborative manner, improving both research quality and adaptability.
Breaking down the framework
① Critic / Generator: two disentangled roles that also swap positions
- Critic Agent (the judge, exploration-leaning): evaluates the quality of the current outline, identifies weaknesses and gaps, and accordingly refines the search queries—it is responsible for "this version isn't good enough yet, we should go search for these".
- Generator Agent (the producer, exploitation-leaning): based on the refined queries, fetches new results and revises the outline—it is responsible for "absorbing the retrieved material into the structure".
- Role swapping produces supervision signals: the Generator is later reused as a scoring agent, evaluating the Critic's output in turn and generating quality signals. This "you grade me, I grade you" adversarial-collaborative closed loop lets both sides supply training/filtering signals to each other without human annotation—this is where the Co (collaboration) in "DisCo" lands.
The two alternate and iterate, and the outline gradually converges through "being critiqued → supplementary retrieval → revision → being critiqued again". It is finally handed to the Report Writer, which synthesizes a long report from the finalized outline; optionally, a Rendering Agent can then render the report into a visual poster.
② The meta-optimization shell and the policy bank: auto-tuning agents via code generation
AgentDisCo's second layer of innovation lies in "who designs this agent configuration". Instead of manually tuning prompts and workflows, it uses code-generation agents (including the Claude-Code family) as a meta-optimization shell, systematically exploring different agent configurations and distilling the designs that work into a policy bank—a structured, reusable "repository of design strategies". Subsequent tasks can directly draw validated strategies from the policy bank, enabling continuous self-refinement with minimal human intervention. This turns "tuning the agent" itself into an automatable, accumulable process.
③ GALA: a new benchmark that "mines" research needs from user browsing history
To make evaluation closer to real usage, the authors propose the GALA (General AI Life Assistants) benchmark: mining latent research interests from a user's historical browsing behavior and constructing research tasks accordingly—compared with "an examiner inventing questions out of thin air", GALA better reflects "what real users actually want to deep-research". The paper also builds an end-to-end product demo on top of this, AutoResearch Your Interest.
Evaluation and performance (per the original paper)
AgentDisCo uses Gemini-2.5-Pro as its backbone and is evaluated on three report-quality-oriented deep research benchmarks:
| Benchmark | Focus |
|---|---|
| DeepResearchBench | Comprehensive deep research capability |
| DeepConsult | Consulting/advisory-style long-report quality |
| DeepResearchGym | Environment-grounded evaluation of deep research tasks |
The paper reports that its performance is on par with, or even better than, leading closed-source deep research systems (exact scores per the tables in the arXiv original). Note that these three benchmarks measure report-writing quality (structure, coverage, argumentation, citations), which differs in focus from browsing benchmarks like BrowseComp/GAIA that test "finding deeply buried information"—AgentDisCo sits toward the "synthesizing the final report" end of deep research work.
Where it sits in the Deep Research lineage
- vs STORM / Co-STORM: both lean toward "organizing long reports", but STORM relies on multi-perspective questioning for pre-writing, whereas AgentDisCo explicitly disentangles exploration/exploitation and adds a layer of code-generated meta-optimization to auto-tune agents, achieving a higher degree of automation.
- vs training-side deep research (Tongyi / REDSearcher): Tongyi DeepResearch and REDSearcher build a long-horizon search model from the training side; AgentDisCo does not train the backbone but instead performs agent orchestration and self-optimization on top of a strong general-purpose model (Gemini-2.5-Pro), complementing that line of work.
- Multi-agent collaboration: its Critic/Generator mutual-critique and role-swapping mechanism is a concrete instance of multi-agent ideas applied to the deep research setting. For the overall positioning, see the Deep Research overview.
References
- Jiarui Jin, Zexuan Yan, Shijian Wang, Wenxiang Jiao, Yuan Lu. AgentDisCo: Towards Disentanglement and Collaboration in Open-ended Deep Research Agents. arXiv:2605.11732, 2026-05 (v2 2026-06). https://arxiv.org/abs/2605.11732
- Evaluation benchmarks: DeepResearchBench, DeepConsult, DeepResearchGym; new benchmark GALA (General AI Life Assistants)