More intelligence.
Less overhead.
Efficiency in LLM Multi-Agent Systems: A Survey
When is collaboration worth its cost? A unified view of how we build, run, and optimize efficient multi-agent systems.
1 National University of Singapore 2 UC Berkeley
3 The Hong Kong Polytechnic University * Equal contribution
Drag a node. Trace its connections. Reshape the team.
Not just fewer agents. Not just fewer tokens.
Collaboration has a cost.
When does it pay off?
Multi-agent systems expand what models can search, verify, and solve. But repeated inference, growing message histories, and coordination can outweigh those gains.
This survey connects the benefit of collaboration to its end-to-end resource cost. It first asks when collaboration is worth invoking, then organizes efficiency methods by the object they control.
Decomposition
Can the task be separated without excessive dependencies?
Information
Does collaboration acquire, preserve, or expose useful evidence?
Escalation
Does the expected benefit justify activating another agent?
The comparison matters: use a capable single controller when feasible, or a simpler multi-agent workflow when multiple actors are intrinsic to the task.
Go inside
the architecture.
Explore the paper’s original figures.
Hover to inspect. Select a region to focus.
A field, connected.
Explore survey references and subsequent additions.
Topic tags indicate relevance to the survey.
Explore the gains.
Keep the context.
Move through an optimization pipeline.
Then compare what different resources reveal.
Base system
The initial system, before the reported optimization pipeline.
Reconstructed from the survey’s MASS results. Stages are cumulative and ordered; increments are not independent causal effects. Read the study ↗
Common normalized scale: 0–150% of baseline
F1 improves from 0.83 → 0.87. Fewer calls and tokens do not guarantee lower latency.
Find the study ↗ Values reproduced from the survey’s synthesis: 962 → 643 calls, 256k → 236k tokens, and 159.4 → 201.5 seconds. Within-study comparison only.Count the whole lifecycle.
Reusable prompts, workflows, and policies move expenditure to design time. Their value depends on how many future requests can reuse them.
- Match model, tool, and information access.
- Measure realized resources, including retries.
- Report quality alongside the relevant cost.
- Include search, training, and re-optimization.
Open questions.
Practical directions.
01 Comparable evaluation +
Can shared budgets, matched baselines, and quality–resource curves make efficiency claims comparable across tasks and model pools?
02 Adaptive control +
Can a controller predict when another collaborative action is worth its incremental cost, then continue, escalate, or stop as evidence arrives?
03 Transfer under drift +
How long does an optimized configuration retain its advantage as models, tools, prices, and workloads change?
04 Causal attribution +
Does the gain come from the graph, role prompts, model access, information routing, or simply additional samples? Matched trace interventions can separate these effects.
05 End-to-end cost +
Can logical work, state reuse, and scheduling be optimized jointly while including tool execution, recovery, defenses, and persistent memory?
Build on the survey.
If you find this survey useful, please cite our work.
@article{202609.1639,
doi = {10.20944/preprints202609.1639.v1},
url = {https://doi.org/10.20944/preprints202609.1639.v1},
year = 2026,
month = {September},
publisher = {Preprints},
author = {Qi Li and Lexiang Xiong and Haiquan Lu and Wenjie Qu and Xingyi Yang and Jiaheng Zhang and Xinchao Wang},
title = {Efficiency in LLM Multi-Agent Systems: A Survey},
journal = {Preprints}
}