Multi-agent research systems

A multi-agent research system is a research process run by several language-model agents: a coordinator that plans and assigns work, parallel workers that search and read, and often verifiers that check what the workers report. In many designs the agents share a board of open questions, and sources raise new questions as the run proceeds. The design problem this article covers is how such a system should bound that growth without cutting off a distant question that matters, how it should rank the new questions, and whether the coordinator should vet what workers propose. The published evidence on these questions is thin and mostly drawn from neighbouring tasks, so the answers below are largely design options with partial measurements.

History

Shared-board designs are older than language models. The builders of the Hearsay-II speech-understanding system reported that the uniform, data-directed control structure was easy to understand and modify, and that their solution to the problem of deciding where to focus attention survived changes to the knowledge sources (Lesser & Erman, 1977). A 2026 practitioner blog says that Hearsay-II’s scheduler estimated how much each candidate action would improve the solution and ran the most promising one, and that letting every specialist fire without constraint exhausted compute (Pan, 2026b).

Focused web crawling met a similar problem from the other side: a crawler keeps a frontier of unvisited pages and must decide which to fetch next. A 2000 context-graph crawler addressed credit assignment to pages along a crawl path, and its authors reported significant efficiency gains over standard focused crawling (Diligenti et al., 2000). Chakrabarti and colleagues later proposed an apprentice crawler that scores each link on its own features Chakrabarti et al. (2002).

In 2025 Anthropic described a system in which a lead agent decomposes the query and decides whether more subagents are needed; subagents are not described as adding tasks themselves (Hadfield et al., 2025). Its early versions over-spawned subagents, which is why effort-scaling rules were later put into the lead agent’s prompts, and the authors report that multi-agent runs use about 15 times the tokens of chat (Hadfield et al., 2025).

Core concepts

Roles and the shared board

A coordinator decomposes a goal into questions, workers answer them in parallel, and verifiers check the answers. Whether workers may add questions to the shared list is a design choice. In Anthropic’s system only the lead agent decides on further subagents (Hadfield et al., 2025). In LATTE, a design for language-agent teams, workers propose changes to a shared task graph and a lead accepts or rejects them (Mieczkowski et al., 2026). RepoMAS, which cites LATTE, logs new requirements and failures found during a run as structured issues that revise the task; its abstract describes no central vetting (Song et al., 2026a). AgentCue studies decentralised coordination in which workers choose their own sub-tasks on a shared board, and reports information overload as the main problem (Zhu et al., 2026).

Bounding growth

Published systems bound the growth of questions in four main ways.

• Depth and breadth limits. Static-DRA sets a depth parameter and a breadth parameter that shrinks at each deeper level; its text says the breadth is halved, though its algorithm subtracts 2. The rule was not ablated and no cost was measured (Prateek, 2025). FlashResearch caps adaptive breadth at 6 and deepens a branch only when the gain per unit of latency passes a threshold (Nie et al., 2025).
• Pruning. FlashResearch also prunes low-yield tree nodes and their descendants early once the research goal is satisfied, scoring them with a language model for goal satisfaction and quality (Nie et al., 2025).
• Effort rules and stop signals. Beyond Anthropic’s prompt rules, a novelty-and-coverage controller that decides when deep research agents stop cut search calls by an average of 14 and raised accuracy by 3% on average on BrowseComp-Plus (Soudani et al., 2026). A programmatic exhaustion gate that stops unproductive agent search saved up to 39% of tokens with no reported loss of performance across three question-answering domains (Kausik et al., 2026).
• Blackboard control. In a language-model blackboard system, runs without a control unit used about 2.9 to 4.5 times the tokens of runs with it, with accuracy within about 0.5 points, (Han & Zhang, 2025).

The quality differences are small. In the FlashResearch ablation that removed planning and orchestration (10-minute budget), quality was 83.90 versus 82.69, the number of nodes explored was 98.43 versus 68.00, and cost was $0.6811 versus $0.5629 per question, for the full system and the ablation respectively; the quality gap is small and pruning is not isolated from the other removed components (Nie et al., 2025).

Ranking added questions

Several fields rank items in a frontier. In focused crawling the baseline gives an unvisited page the priority of its parent (Chakrabarti et al., 2002), and a classifier strategy of the same kind did not perform well against query-based strategies (Pirkola & Talvensaari, 2009). The alternative is to score each new item on its own features. In the apprentice crawler this found crawl paths that the baseline never expanded (Chakrabarti et al., 2002) and cut the loss rate by 30-90 percent compared with the baseline (Chakrabarti et al., 2002). A page-level variant, which scores blocks of a page separately so that the crawler can pass through a page that is irrelevant as a whole, reports a higher harvest rate than best-first and breadth-first crawling (Luo & WanLi, 2007).

A second family of methods estimates how far an item is from a payoff. The context-graph crawler regresses from page text to the estimated link distance to a relevant page, and Chakrabarti and colleagues report that Diligenti and colleagues showed significant improvements in harvest rate on topics where harvest rate was low (4-6%) (Chakrabarti et al., 2002). A tunnelling crawler with depth and relatedness parameters did better later in a crawl than the basic crawler but worse early, in a preliminary study on one topic (Tsoi et al., 2003).

For language agents choosing what to ask next, scoring by expected information gain beat purely prompting-based design generation and other adaptive strategies on 20 Questions and preference-inference tasks (Choudhury et al., 2025). Uncertainty of Thoughts, which rewards questions by information gain, improved task completion by 38.1% on average over direct prompting across several models. It limited simulation to three steps because of computation and API budgets, and used about 9.2k GPT-4 tokens versus 4.7k for the pruned version (Hu et al., 2024).

Rating priority with a language model

A coordinator that sets priority is in effect a judge (see LLM-as-a-judge). Single-response Likert ratings from language models are unstable under sampling and compress near the top of the scale, with frequent ties (Girrbach et al., 2025). With a fine-grained ordinal scale, pointwise scoring closes much of the gap to listwise ranking, and the gap is often not statistically significant, although listwise ranking is the research community’s prevailing view as the better method (Godfrey et al., 2025). Pointwise scores and pairwise preferences from the same judge often contradict each other; in one study widening the score scale from 5 to 100 points reduced the conflicts (Wang et al., 2025). A judge that sees a whole list is order-sensitive: in decoder-only rerankers the input order affects scores, preferences and rankings (Bito et al., 2026). Shuffling the list order and aggregating the rankings improved ranking scores by up to 7-18% for GPT-3.5 and 8-16% for LLaMA v2 70B over a single pass (Tang et al., 2023), and averaging sampled judge scores correlated better with humans than greedy decoding or median or majority voting (Yamauchi et al., 2025).

Human practice

Human evidence synthesis is a source of tested rules. In scoping studies the search strategy is refined iteratively from what early results show, while the research question is stated up front (Levac et al., 2010). The Cochrane handbook requires changes to a protocol after a review starts to be documented and reported, and asks for sensitivity analyses of the effect of deviations where possible (Cochrane, 2023). Library guidance says to fix any stopping rule, such as screening a fixed number of results or stopping after a run of irrelevant results, before the search begins and to document it (Radboud University Library, 2026). Stopping rules for screening records are hard to get right: of 15 methods evaluated on 81 reviews, only one reliably reached the recall target, and it stopped conservatively (Repke et al., 2026).

Questions, controversies, and new developments

Do growth bounds cut off distant questions?

A bound that stops too early costs recall. Deep research agents often stop after collecting limited evidence, giving uneven depth, and removing agent-level termination criteria in one enterprise system lowered quality (HAA 82.09 to 73.67, coverage 4.31 to 4.02) while also lowering tool calls, which the authors attribute to premature termination; the result is limited to one task setting (Choubey et al., 2026). A cap on search steps costs the questions whose evidence arrives late: one analysis estimates that a 20-episode cap concedes at most 14% of accuracy, an estimate that assumes the answer is unchanged under early stopping, so hard truncation may lose more (Liu et al., 2026c). Both measure step or termination limits rather than caps on added questions. Against this, the novelty and exhaustion signals above saved calls or tokens with small accuracy changes in their own tests (Soudani et al., 2026; Kausik et al., 2026).

Does inherited priority cut off far-away relevant items?

In crawling the evidence is mixed. Relevance to a topic is reported to drop within a few links of a relevant start page (Chakrabarti et al., 2002). Per-link scoring was supported by later studies: a shared-baseline comparison found every crawler variant performed best when scoring links with both page content and anchor text (Batsakis et al., 2009), an adaptive crawler retrieved more target forms than fixed-focus crawlers in hidden-web search (Barbosa & Freire, 2007), and the choice of classifier matters for guiding a crawler, with Naive Bayes weak (Pant & Srinivasan, 2005). Some evidence points the other way. An uninformed breadth-first ordering was reported to be a good strategy for finding high-quality pages early, though quality declines as the crawl progresses (Najork & Wiener, 2001). A per-page classifier strategy performed poorly against query-based strategies using link-context words, and training a classifier was named as harder than stating the topic as a query (Pirkola & Talvensaari, 2009). None of these sources tests a cap with decay against per-item scoring directly.

Should the coordinator arbitrate proposed questions?

The comparisons are indirect. In multi-robot planning, the decentralised design had the lowest success and highest token use, a single central planner scaled worse than a hybrid as agents were added, and the hybrid beat both pure designs (Chen et al., 2024a). In LATTE, leader-worker teams reached 70% accuracy versus 74% for decentralised teams and 80% for the propose-and-review design (a test-suite pass rate averaged over three tasks) (Mieczkowski et al., 2026). Architectures without central verification propagated errors more than those with central coordination, and whether centralised coordination helps depends on task fit, with mismatched coordination degrading performance (Kim et al., 2025c).

Central control has reported costs. Anthropic reports bottlenecks from synchronous central control (Hadfield et al., 2025), and the LATTE paper names bottleneck and single-point-of-failure costs of centralisation (Mieczkowski et al., 2026). In a paired experiment on an open-ended task, flat teams produced higher-quality reports than hierarchical teams with a manager who could send work back (Agachan et al., 2026); manager review is reported to help only when the manager can verify the work, and with feedback alone to hurt quality (Agachan et al., 2026). These studies concern robots, code and report writing, not question lists, and most are preprints.

Can a language model merge duplicate questions?

If an arbiter merges duplicates, a false merge deletes a distinct line of inquiry. A fine-tuned LLaMA-7B reached F1 of 84.9% on duplicate-question identification on Quora Question Pairs, (Han et al., 2025a). In entity matching, language models fail on names written in different scripts while rule-based matching over-matches, so the two fail in complementary ways (Smith et al., 2026).

Are three priority levels enough?

Evidence on scale granularity is mixed and comes from relevance and quality grading. In a controlled comparison of three grading scales, a 0-5 scale gave the strongest human-model agreement and the choice of scale substantially shifted agreement; wider scales were not always better (Li et al., 2026a). A 2026 workshop paper found that finer scales often gave higher agreement with human labels in argument-quality assessment, and reports that for human raters fewer possible scores can make decisions clearer and more consistent (Marchetti et al., 2026). Lee and colleagues study evaluator self-consistency and consistency across scoring scales and criterion granularity (Lee et al., 2024). Rankings for triage and allocation should be checked for run-to-run variance and transitivity, and three leading models differed on both (Pokharel et al., 2026). A multi-agent system for user-story prioritisation agreed only moderately with expert rankings, best on mid and low priority items, and varied across runs (Sami et al., 2025).

Broader connections

The coordinator’s priority ratings face the same validity problems as any language-model grader, treated in LLM-as-a-judge, and checking a multi-agent run for repeatability and cost uses the tooling described in Evaluation tooling for AI agents. Packaging coordinator and worker roles as reusable instruction sets is the subject of Agent skills. Deciding which sub-question to pursue, and when to stop, is a search through a problem space in the sense of Problem solving. The frontier-priority problem is shared with focused crawling, and the stopping and protocol rules with human evidence synthesis, as described above.

Further reading

• Chakrabarti et al. (2002): a measured comparison of parent-inherited and per-item priority in a frontier.
• Nie et al. (2025): a tree-structured research agent with breadth caps, pruning and a quality-cost ablation.
• Mieczkowski et al. (2026): the closest design to propose-then-review for task changes, compared with leader-worker and decentralised teams.
• Hadfield et al. (2025): an engineering account of a production lead-and-subagents research system, including over-spawning and token costs.
• Levac et al. (2010): how scoping studies keep the question fixed while the search strategy evolves.
• Liu et al. (2026c): what a step cap costs a search agent.