From 26% to 71% with the Same Model: How an AI Swarm Wins

In academic research, "multi-Agent" can mean role-based collaboration (Li et al., 2023), multi-round debate (Du et al., 2024), or emergent group behavior (Chen et al., 2024). In industrial systems, a "swarm" more often means a main Agent scheduling multiple workers (Anthropic, 2025). Similar names are used for systems evaluated by different standards: accuracy and cost, robustness, or whether Agents can autonomously develop divisions of labor and relationships.

This article therefore asks only two direct questions: can a swarm fully decompose a task, complete the parts independently, and merge them reliably; and can that division of labor and connectivity gradually emerge from the Agents themselves? The first determines whether a swarm can deliver a reliable artifact. The second determines whether it has emergent capability.

For large tasks, the ideal is neither to put more Agents into the same group chat nor to let a "super-manager" decide everything for everyone. A better arrangement is to fully decompose a large task, give each Agent a clearly bounded part, ensure that all parts jointly cover the original task, and merge them again through reliable interfaces.

Figure 1. The core comparison framework. The EvoX swarm lets multiple Agents complete one part each and merge the results. The Sub-Agent mode relies on a main LLM to decompose and summarize. The single-context mode puts the entire task into one context. The upper layer shows the target form of a complete swarm. Experiment 1 directly validates complete coverage, independent contexts, and programmatic aggregation; autonomous decomposition, autonomous task claiming, and information exchange between Agents still require validation in later experiments.

We ran two experiments corresponding to the two research questions. Experiment 1 asks whether a swarm can win when complete division of labor is already in place. Experiment 2 asks whether Agents can start selecting partners and forming a structure when the entire organizational relationship is not specified in advance.

Experiment 1: The Swarm's Advantage Is More Than "More Agents"

The experiment used a fixed set of 563 questions: 100 logic questions, 250 standard mathematics questions, 63 competition mathematics questions, and 150 physics questions, sourced respectively from ZebraLogic, MATH, AIME/HMMT, and UGPhysics. All three modes used the same Claude Haiku 4.5. The question set and grading standard were held constant; only task decomposition, execution, and aggregation differed.

Figure 2. Composition of the Experiment 1 question set: 563 questions spanning logic, standard mathematics, competition mathematics, and physics.

The first mode was the EvoX swarm. The target task was fully split into as many atomic tasks as possible. Each question was assigned to an independent Agent. Each Agent completed one part at a time and recorded the result in a designated location. A program designed by the main Agent then collected the answers by question number. No second model rewrote the results, and no central role decided which answers should be kept.

The second was the Sub-Agent mode, which simulates the main-Agent scheduling mechanism common in products such as Claude Code and Codex. A main coordinating LLM first saw the complete question list and divided it into a series of subtasks, some of which could be split further. Sub-agents solved questions, organized reports, and returned them to the main LLM through continuous sessions. The main coordinating LLM then combined 30 reports into the final answer table.

The third was the single-context mode: one Agent directly solved the entire target task and received all questions at once.

The results formed three very clear tiers:

• EvoX swarm: 70.69%-70.87%
• Sub-Agent mode: 38.54%
• Single context: 26.29%

Figure 3. With the same model and question set, the three organizational modes produced three distinct levels of final performance. One network request timed out in the EvoX swarm, so the upper and lower bounds are reported using the fixed denominator of 563 questions.

This was not simply a case of "calling the model more times, so performance naturally improves." The Sub-Agent mode also called many sub-agents, and was even the most expensive of the three systems by observed cost. It still fell far short of the EvoX swarm.

Existing research likewise shows that the benefits of multiple Agents do not appear automatically as the number of Agents increases. Naive chaining can amplify hallucinations, while role definitions, coordination methods, and result verification can themselves become new failure points (Hong et al., 2024; Cemri et al., 2025).

The swarm won for three reasons.

First, it decomposed the task more effectively. By encouraging the task to be split into as many atomic parts as possible, it made each subtask as small as possible. Every question had a clear handler and output location. Each atomic task could be tracked and solved more quickly.

Second, execution was isolated. Each Agent faced a clearly bounded local problem. It did not need to switch repeatedly inside a huge context, nor was it continuously distracted by the accumulated content of dozens of earlier tasks. Research on long contexts has also found that models cannot reliably use information in the middle of a context. That finding is directionally consistent with this experiment, but this experiment did not independently manipulate question position, so the gap cannot simply be attributed to "lost in the middle" (Liu et al., 2024).

Third, aggregation no longer depended on re-understanding. A predefined program collected atomic-task results from their agreed locations. Correct answers did not have to pass through another LLM for paraphrasing, compression, or selection.

The core of a swarm is therefore not just high parallelism. It is the transformation of a complex task into parts with clear boundaries that can be completed independently and merged reliably.

However, 38.54% was only the Sub-Agent mode's final delivery score. To determine whether it lost because the sub-agents could not solve the questions or because answers were lost during transmission and aggregation, we traced the intermediate results left by the 30 subtasks and rechecked them question by question.

The Sub-Agent Mode Lost 166 Correct Answers. What Did the EvoX Swarm Get Right?

An unexpected pattern appeared. Many questions had initially been solved correctly: sub-agents had produced correct answers for 166 questions that did not safely reach the final result.

Among the 563 questions, 373 were judged correct in the intermediate results. After reports were transmitted and the main coordinating LLM produced its final synthesis, only 217 correct answers remained in the delivered output. Of these, 207 stayed correct from intermediate result to final result, while 10 changed from wrong to right. More strikingly, 166 answers that had been correct at one point became wrong or missing in the final delivery. The retention rate for intermediate correct answers was only 55.50%.

Figure 4. Of the 373 correct intermediate answers in the Sub-Agent mode, 166 were not preserved in the final result.

This chain resembles a game of telephone. Every additional layer of natural-language paraphrasing creates another opportunity for omission, compression, format drift, or erroneous overwriting. Once one link fails, the later stages can easily compound the damage.

The 166 questions should not all be blamed on the final aggregation step. Context accumulation across continuous sessions, the way sub-agents wrote reports, and the way the main coordinator synthesized them jointly formed the information-loss chain. Recent research on multi-Agent failures likewise distributes the problem across system design, Agent-to-Agent alignment, verification, and termination (Cemri et al., 2025).

The EvoX swarm did the opposite: it did not make correct answers go through another round of "being understood." Each result was written to a fixed location and collected by a program according to question number. The Agent solved the question; the program merged the results. No other LLM had to read 30 reports and decide again which answers should be kept.

As a result, the same model answered 217 questions correctly in the Sub-Agent mode and 398 correctly in the EvoX swarm. The gap came not only from finer task decomposition, but also from safely transporting answers that were already correct to the finish line.

The EvoX Swarm Has Succeeded, but This Is Not the End of the Story

Experiment 1 validated the execution layer of a swarm: when complete division of labor already exists, each Agent completes its own part, and the system can merge the results without loss, a swarm can significantly unlock the capability of the same model.

But the complete decomposition was still guaranteed by a script. Real-world tasks are rarely as tidy as 563 independent questions. Some parts depend on one another, some require different specialties, some are created during execution, and some fail, conflict, or duplicate one another.

This leaves a harder and more interesting question: can a swarm move from "the script provides complete division of labor" to "Agents discover specialties, claim tasks, choose partners, and adjust the organization themselves"?

Experiment 2 began with the smallest version of that action: choosing a partner.

Experiment 2: A Free Swarm Develops a Coherent Organizational Form

Experiment 2 first asked 24 identically configured Agents to solve questions. After completing each question, an Agent summarized what it had learned about that question type and wrote the experience into persistent memory. In the experiment, these experiences were recorded as Genes, represented by a skill field. The more experience an Agent accumulated in a category, the more likely it was to choose that category in the next round. After eight rounds, the system used each Agent's actual choices and results to form two pieces of identity information: specialization profile and accuracy.

Take agent-8 as an example. It started exactly like the other Agents, but across eight rounds it chose physics five times, standard mathematics twice, and competition mathematics once. It ultimately developed a specialization profile of "physics 0.625, mathematics 0.25, competition mathematics 0.125," with an accuracy of 0.25. This physics-oriented identity was not a role written in advance. It was a personal history formed through repeated question choices and accumulated memory.

The 24 Agents then started from the same "round-table" relationship network. Each Agent was connected to the two Agents on either side, giving it four relationships in total. The experiment randomly broke some connections and asked the affected Agent to choose a new partner from candidates it was not already connected to.

The three experimental groups had the same initial Agents, broken edges, and candidate sets. The only difference was what an Agent could see when choosing a partner:

• connection relationships only;
• connection relationships plus specialization profile and accuracy;
• no preference judgment, with the system reconnecting randomly as a control.

The same Agent used entirely different selection logic simply because the visible information changed.

Figure 5. The same agent-8 chose different partners depending only on the information it could see: a friend of a friend when it saw relationships alone, and a complementary partner when it could see task information.

In one actual decision, agent-8 chose agent-7 when it could see only social relationships because the two shared four existing connections, a classic "friend of a friend" choice. In network science, this mechanism is called triadic closure and is a classic mechanism through which human relationship networks form clusters (Watts & Strogatz, 1998; Papachristou & Yuan, 2025). agent-7's accuracy was only 0.25, but agent-8 could not see that at the time.

When specialization profile and accuracy became visible, the same agent-8 instead chose agent-2, which had no common neighbors. agent-2 had an accuracy of 0.75, and its specialization profile complemented agent-8's. The model's verbal explanation of candidate information was not perfectly accurate every time, but the overall direction of its choices changed consistently: when it saw relationships alone, it looked for familiar connections; when it saw task information, it began choosing partners.

This change was not limited to an individual example. With social information alone, "choose a friend of a friend" was the strongest preference, at +2.28. After task information was added, that preference fell to +0.19 and was replaced by "choose someone with high accuracy" at +1.88 and "choose someone with a similar specialization" at +1.47. Choosing people similar to oneself is commonly called homophily in social-network research (McPherson et al., 2001).

The Same Group of Agents Developed Two Social Forms

As many local choices accumulated, the entire network changed shape.

When Agents could see only social information, they repeatedly connected to friends of friends, and the network retained clear clusters. Its overall clustering coefficient was approximately 0.53. After specialization profile and accuracy became visible, Agents were more willing to cross existing clusters to connect with high-performing peers. The clustering coefficient fell to 0.28, close to the random-rewiring control at 0.27, and more prominent hub nodes appeared. The round-table starting point, rewiring process, and clustering metric came from the classic small-world network framework. A strict judgment about whether this is a "small world" would also require comparing path length with random and regular-network baselines (Watts & Strogatz, 1998; Telesford et al., 2011).

Figure 6. With social information alone, cluster structure was preserved. After task information was added, the network moved toward lower clustering and greater centralization.

Figure 7. The same group of Agents formed different organizations simply because the visible information differed. Lines represent rewiring choices by Agents, not message flow or task handoffs. The network shown is an illustrative example from seed 5004; the overall conclusion is based on multiple observations.

One possible next step is to let these connections carry real collaboration. Agents could exchange Genes along the relationships they form, hand off tasks, find complementary specialists, and look for replacements after failure. At that point, the network would no longer be only a shape observed in an experiment. It would become a task-routing and experience-sharing system that Agents actually use.

EvoX: Toward a Self-Organizing Swarm

Taken together, the two experiments make EvoX's next step clear.

The first layer is a reliable execution substrate for swarms: every part has a stable ID; task coverage can be checked; results return through structured interfaces; duplicates and gaps can be detected; failures can be retried; and final merging is handled by a program whenever possible. Experiment 1 shows that these seemingly simple engineering constraints directly determine whether a swarm can carry individual capability all the way to delivery.

The second layer is an adaptive self-organization mechanism. Agents should be able to expose and update their specialties, propose task decompositions, claim parts, choose collaborators, and reassign work when they discover dependencies or conflicts. Experiment 2 suggests that organization does not appear from nowhere. The information exposed to Agents determines which connections and structures the system rewards.

This also means that free division of labor does not mean having no rules. The more autonomy Agents are expected to have, the more reliable the underlying protocols need to be. Freedom belongs in decisions about who does what, how parts are combined, and when the structure should change. Task coverage, result formats, and error handling should not be left entirely to chance.

A real AI swarm is not just more Agents talking at the same time.

It lets each individual complete its own part, ensures that all parts jointly cover the full task, and helps the right Agents find one another at the right time.

Experiment 1 showed the potential of this kind of swarm. Experiment 2 showed that it may also begin to grow its own organization.

References

Multi-Agent Collaboration and Evaluation

• Li, G., Hammoud, H. A. A. K., Itani, H., Khizbullin, D., & Ghanem, B. (2023). CAMEL: Communicative Agents for “Mind” Exploration of Large Language Model Society. arXiv:2303.17760.
• Du, Y., Li, S., Torralba, A., Tenenbaum, J. B., & Mordatch, I. (2024). Improving Factuality and Reasoning in Language Models through Multiagent Debate. ICML 2024, 11733–11763.
• Chen, W., et al. (2024). AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors. ICLR 2024.
• Wu, Q., et al. (2024). AutoGen: Enabling Next-Gen LLM Applications through Multi-Agent Conversation. COLM 2024.
• Hong, S., et al. (2024). MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework. ICLR 2024.
• Cemri, M., et al. (2025). Why Do Multi-Agent LLM Systems Fail? arXiv:2503.13657.
• Zhu, K., et al. (2025). MultiAgentBench: Evaluating the Collaboration and Competition of LLM Agents. ACL 2025, 8580–8622. doi: 10.18653/v1/2025.acl-long.421.

Long Context and Experimental Methods

• Liu, N. F., et al. (2024). Lost in the Middle: How Language Models Use Long Contexts. Transactions of the Association for Computational Linguistics, 12, 157–173. doi: 10.1162/tacl_a_00638.
• Liang, P., et al. (2023). Holistic Evaluation of Language Models. Transactions on Machine Learning Research.
• AWS. (n.d.). Claude Haiku 4.5 model card.

Network Science and LLM Network Formation

• Watts, D. J., & Strogatz, S. H. (1998). Collective dynamics of “small-world” networks. Nature, 393, 440–442. doi: 10.1038/30918.
• McPherson, M., Smith-Lovin, L., & Cook, J. M. (2001). Birds of a Feather: Homophily in Social Networks. Annual Review of Sociology, 27, 415–444. doi: 10.1146/annurev.soc.27.1.415.
• Telesford, Q. K., Joyce, K. E., Hayasaka, S., Burdette, J. H., & Laurienti, J. P. (2011). The Ubiquity of Small-World Networks. Brain Connectivity, 1(5), 367–375. doi: 10.1089/brain.2011.0038.
• Papachristou, M., & Yuan, Y. (2025). Network formation and dynamics among multi-LLMs. PNAS Nexus, 4(12), pgaf317. doi: 10.1093/pnasnexus/pgaf317.
• Anthropic. (2025). How we built our multi-agent research system.