Multi-Agent Systems · Cost-Aware Architecture

The decomposer is
the bottleneck.

Researchers at East China Normal University factorized a hierarchical search agent into three separated roles, ran controlled capacity sweeps across each one, and found that model size barely matters where most builders put the budget. Scaling the delegation layer lifts accuracy by roughly 11 points. Scaling the execution layer moves it less than 3. The sub-agents doing retrieval can be kept small.

First surfaced in Tandemly Briefing — 2026-08-07.

Core concept
Capacity sensitivity is asymmetric in hierarchical search agents. Decomposition is the bottleneck: a strong delegation backbone transforms how search proceeds. Execution sub-agents can be small models distilled from frontier trajectories without meaningful accuracy loss.
scroll to explore

Bigger is better everywhere
isn't a strategy.

Multi-agent search systems are expensive. Most builders respond by scaling uniformly, using the same tier of model throughout. That turns out to be the wrong question.

When an agent tackles a complex question, it rarely handles it in one pass. The more robust pattern is hierarchical: a coordinator breaks the question into sub-problems, hands each sub-problem to a retrieval agent, and a writer pulls the answers together into a final response. This architecture is common. The question of where to put the compute budget inside it is not.

The field's default answer is something like: bigger is better everywhere. Use a frontier model for the coordinator and frontier models for the sub-agents, and assume the bottleneck is overall capability. This produces high-quality systems that are also expensive, and it tends to discourage experimentation because every component is already maxed out.

What the field hasn't had is a clean separation of roles followed by systematic variation of model size at each role. Without that, builders cannot know whether they're over-spending on execution and under-spending on decomposition, or vice versa. They're flying blind on where the dollar actually buys accuracy.

The question this paper asks

In a hierarchical search agent with cleanly separated roles, where does scaling model capacity actually move the needle on accuracy? Is the bottleneck in how the query gets broken down, in how sub-agents retrieve information, or somewhere else entirely?

Three roles, each varied
independently.

The paper builds a hierarchical multi-agent search system with three named roles, then holds two fixed while scaling the third. The experimental design separates the question of where capacity matters from the question of how much capacity you have in total.

The system architecture has three separated components. A delegation agent, sometimes called a Shepherd, receives the original query and decomposes it into sub-questions. It maintains global context across the search process and directs the population of retrieval agents. A set of execution sub-agents handle search and retrieval: each takes one sub-question, looks for relevant information, and returns it. A fixed answer writer synthesizes everything into a final response. The writer does not vary across experiments.

The core move is capacity isolation. To measure the contribution of each role, the researchers vary the model tier at one role while holding the others constant, then measure exact-match accuracy across five multi-hop question-answering benchmarks. Multi-hop tasks are a good test because they require genuine decomposition: the right sub-questions depend on understanding how the problem connects across multiple retrieval steps.

Delegation
Coordinator
Decomposition
Receives the original complex query. Breaks it into sub-questions. Directs execution agents. Maintains global context across the entire search process. This is where the paper finds most of the capacity leverage.
Execution
Sub-agents
Retrieval
Each sub-agent receives one sub-question and performs search and retrieval. Scaling this role produces much smaller accuracy gains than scaling delegation. The paper shows a 1.7B distilled model can match a frontier sub-agent here.
Writer
Fixed
Synthesis
Synthesizes retrieved information into a final answer. Held constant across experiments to isolate the contribution of the other two roles. Not the subject of the capacity sweep.

Alongside the capacity sweep, the researchers train a compact executor through quality-filtered trajectory distillation. The idea is straightforward: collect successful search trajectories from a frontier sub-agent, filter them for quality, and use them to train a much smaller 1.7-billion-parameter model. If the distilled model's search behavior is close enough to the frontier model's, it can replace the expensive sub-agents at a fraction of the cost.

Why multi-hop QA is the right test

Multi-hop question answering requires the kind of genuine decomposition hierarchical agents are built for. Answering "Who directed the film that won the same award as the novel the author of X also wrote?" requires breaking the problem into connected sub-queries, not retrieving a single document. Accuracy on this class of task is a fair measure of whether decomposition quality actually matters.

Eleven points versus
two and a half.

The capacity sweep produces a clean asymmetry. The two roles do not respond to scaling in the same way, and the gap is large enough to have real budget implications.

Scaling delegation backbone
~11
points exact match improvement
Upgrading the delegation backbone to a stronger model produces roughly 11 points of improvement on exact-match accuracy across the five multi-hop benchmarks. A larger decomposer asks better sub-questions, coordinates search more effectively, and maintains global context without losing the thread.
Scaling execution sub-agents
~2.6
points exact match improvement
The same upgrade applied to the execution sub-agents moves accuracy by only about 2.6 points. Execution sub-agents handle well-scoped sub-questions with specific retrieval targets. The task is bounded enough that a smaller model can do it without meaningful accuracy loss.
The distillation result

A 1.7-billion-parameter executor trained via quality-filtered trajectory distillation matches a frontier sub-agent while consuming 37% fewer sub-agent tokens across the five benchmarks. The frontier model's successful search trajectories, filtered for quality before distillation, transfer enough of the retrieval behavior to make the smaller model competitive. This matters because execution sub-agents are the high-volume component: they fire once per sub-question, so cost compounds quickly.

Limitations to keep in mind

The evaluation is on multi-hop question-answering benchmarks, which are a natural fit for hierarchical search. Domains where the decomposition step is simpler, where sub-questions are nearly independent, or where execution involves more complex reasoning than retrieval might show a different capacity profile. The asymmetry is a finding about hierarchical search agents specifically, not a universal law about all multi-agent systems.

Put the strong model
where it earns its keep.

This paper is most useful for teams building hierarchical search agents who have been sizing components intuitively rather than empirically. The findings give a concrete starting point for where to put the budget.

1
For agent architects
If you have a fixed budget and a hierarchical search system, put the strongest model on delegation rather than on execution. The ~11-point versus ~2.6-point gap is large enough to matter in production. A weaker decomposer cannot be compensated for by stronger sub-agents. A stronger decomposer with smaller sub-agents often wins on both cost and accuracy.
2
For teams managing inference cost
Execution sub-agents are the high-volume component of a hierarchical search system. They fire per sub-question, which can mean many calls per user query. If you're paying frontier rates for those calls, the distillation result is directly actionable: quality-filtered trajectory distillation on a 1.7B model can get you to matched accuracy at 37% fewer sub-agent tokens. The distillation approach requires an investment in trajectory collection and filtering upfront, but the token savings compound across every future query.
3
For builders moving from flat to hierarchical architectures
The paper's framing is useful regardless of whether you adopt their specific system. The lesson is that hierarchical multi-agent systems are not uniformly sensitive to capacity. Before defaulting to the largest available model at every layer, run at least one controlled experiment varying model size at each role while holding others fixed. The bottleneck may not be where you expect it.
4
For teams using distillation for agent specialization
Quality filtering before distillation is the key step. Simply collecting trajectories from a frontier model and training on them mixes good search behavior with bad. Filtering for successful, high-quality examples before distillation is what makes the 1.7B model competitive. If you're using trajectory distillation on other agent specialization tasks, the quality filtering step is worth explicit attention and measurement.

Where to go
from here.

Concrete next steps for teams who want to apply the capacity-asymmetry framing to their own systems.

1
Read the paper
Cai, Zhao & Li (2026). Think Big, Search Small: Where Capacity Matters in Hierarchical Search Agents. East China Normal University. arXiv:2607.07548.
2
Audit your own multi-agent system for role separation
Before running a capacity sweep, check whether your system has cleanly separated roles or whether the same component does decomposition and retrieval. Mixed-role components make it harder to isolate which part of the system is the bottleneck. Separating roles cleanly is a prerequisite to the kind of targeted scaling the paper describes.
3
Run a small capacity sweep on your delegation layer
Pick three model tiers for your delegation backbone, hold everything else constant, and measure task accuracy. Even a 20-task held-out set is enough to see whether the asymmetry holds in your domain. If the delegation layer is already maxed out, try the reverse: downgrade sub-agents one tier and measure the accuracy delta. That gives you the execution-sensitivity number for your specific system.
4
Prototype quality-filtered trajectory distillation for execution sub-agents
If execution sub-agents are driving a significant share of your inference cost, collect 500 to 1,000 successful search trajectories from your current frontier sub-agent, filter for high-quality examples (successful task completion, efficient search behavior, no hallucinated retrieval steps), and fine-tune a smaller open-weight model on them. Measure exact-match accuracy and token cost per query against your frontier baseline before deploying.
5
Track cost per role, not just overall cost
Most cost dashboards report aggregate token spend per user query. Adding per-role attribution, total tokens at delegation versus at execution, makes the capacity-asymmetry finding actionable: you can see whether the expensive component is actually the high-leverage one, and adjust budgets accordingly.