A swarm's strength emerges when a complex task is divided into atomic tasks with clear boundaries, each Agent completes its own part independently, and the results are merged reliably. The experiments also show that Agents choose different collaborators based on the information available to them and gradually form different organizational structures.
In academic research, "multi-Agent" systems include role-based collaboration (Li et al., 2023), multi-round debate (Du et al., 2024), and emergent group behavior (Chen et al., 2024). In industry, a "swarm" more often means a lead Agent coordinating multiple workers (Anthropic, 2025). These systems use similar names but are judged by different standards: accuracy, cost, robustness, and whether Agents can form divisions of labor and relationships on their own.
This article examines two questions that directly affect delivery. Can a swarm decompose a task completely, finish each part independently, and merge the results reliably? Can its division of labor and connections gradually be formed by the Agents themselves? The first determines whether a swarm can deliver a reliable artifact. The second determines whether it can become self-organizing.
Large tasks require complete decomposition, clear boundaries, and reliable interfaces. Each Agent completes a local task, all local results collectively cover the original task, and the results are merged through fixed interfaces.
Figure 1 | The central comparison in this article. The EvoX swarm assigns one part to each Agent and merges their results; the Sub-Agent setup asks a lead LLM to decompose and consolidate the work; the single-context setup places the entire task in one context. Experiment 1 validates complete coverage, independent contexts, and programmatic aggregation. Autonomous decomposition, task claiming, and information exchange between Agents remain to be closed in later experiments.
The two experiments address two research questions. Experiment 1 asks whether a swarm can produce a better final result once complete division of labor is in place. Experiment 2 asks whether Agents can choose their own partners and form a structure when the organization is not fully predetermined.
Experiment 1: A Swarm's Advantage Comes from More Than Its Size
The experiment uses a fixed set of 563 questions: 100 logic questions, 250 standard mathematics questions, 63 competition mathematics questions, and 150 physics questions from ZebraLogic, MATH, AIME/HMMT, and UGPhysics. All three setups use Claude Haiku 4.5. The dataset and grading criteria remain the same; only task decomposition, execution, and aggregation differ.
Figure 2 | Composition of the Experiment 1 dataset: 563 questions covering logic, standard mathematics, competition mathematics, and physics.
The first setup is the EvoX swarm. The target task is fully divided into as many atomic tasks as practical. Each question enters an independent Agent, each Agent completes one part at a time, and the answer is written to a designated location. A program designed by the lead Agent collects the answers by question number. No second model rewrites the results, and no central role decides which answers to retain.
The second setup is the Sub-Agent mode. It represents the lead-Agent orchestration common in products such as Claude Code and Codex. A lead coordinator LLM first sees the complete list of questions and divides it into a series of subtasks; those subtasks may be divided further. Sub-Agents solve questions in continuous sessions, prepare reports, and return them to the lead LLM. The lead coordinator then combines 30 reports into the final answer table.
The third setup uses a single context. One Agent solves the entire target task in the same context and produces all answers at once.
The results fall into three clear tiers: the EvoX swarm reaches 70.69%-70.87%, the Sub-Agent setup reaches 38.54%, and the single-context setup reaches 26.29%.
Figure 3 | The same model and dataset produce three tiers of final accuracy. One EvoX swarm question encountered a network timeout, so the range is reported with the fixed denominator of 563 questions.
Calling a model more times does not automatically improve the result. The Sub-Agent setup also calls multiple Agents and, in this experiment, incurred the highest observed cost tier among the three systems. Its final result still remained far below the EvoX swarm.
Prior work also shows that multi-Agent gains depend on organization. Role definition, coordination, validation, and termination can all introduce new failure points (Hong et al., 2024; Cemri et al., 2025).
The swarm's advantage requires three conditions to hold together.
First, the task is divided into atomic units with clear boundaries. Every question has an assigned processor and output location. Each atomic task is traceable and can run independently in parallel.
Second, execution is isolated. Each Agent sees a local problem and does not need to switch repeatedly within a huge context or remain exposed to the accumulated content of dozens of earlier tasks. Long-context research finds that models do not consistently use information in the middle of a context. That finding is consistent with the direction of this experiment, but this experiment did not independently manipulate question position, so the difference cannot be attributed to the "lost in the middle" effect (Liu et al., 2024).
Third, aggregation does not depend on another model reinterpreting the answers. A program collects atomic-task results from agreed locations. Correct answers that have already been produced do not pass through another LLM for paraphrasing, compression, or selection.
The swarm converts a complex task into parts that have clear boundaries, can be completed independently, and can be merged reliably. Parallelism is the outcome; complete coverage and reliable aggregation are the prerequisites.
The Sub-Agent score of 38.54% is only the final delivery score. To determine whether the loss occurred during problem solving or during transmission and aggregation, we traced the intermediate results left by the 30 subtasks and checked them question by question.
The Sub-Agents had answered 373 questions correctly at some point, but many of those correct answers never reached the final result.
The Sub-Agent Pipeline Lost 166 Correct Answers. What Did the EvoX Swarm Do Differently?
Among the intermediate results, 373 of the 563 questions were already correct. After report transmission and the lead coordinator LLM's final synthesis, the delivered result contained only 217 correct answers. Of these, 207 stayed correct from the intermediate stage to the final output, and 10 changed from incorrect to correct. A total of 166 once-correct answers became incorrect or went missing in the final delivery. The retention rate for intermediate correct answers was 55.50%.
Figure 4 | Of the 373 correct intermediate answers in the Sub-Agent setup, 166 were not preserved in the final result.
Every additional natural-language relay introduces another opportunity for omission, compression, format drift, or erroneous replacement. Once an error occurs at one stage, downstream stages can only process the already damaged information.
These 166 questions cannot all be attributed to the final synthesis step. Context accumulation in continuous sessions, the way Sub-Agents wrote their reports, and the way the lead coordinator synthesized them jointly formed the information-loss chain. Recent research likewise locates multi-Agent failures across system design, inter-Agent alignment, validation, and termination (Cemri et al., 2025).
The EvoX swarm avoids making a correct answer undergo another round of interpretation. Each answer is written to a fixed location and collected directly by question number: Agents solve the questions, while a program performs aggregation. Another LLM does not need to read 30 reports and decide again which answers to retain.
With the same model, the Sub-Agent setup ultimately answered 217 questions correctly, while the EvoX swarm answered 398. The difference comes from both task granularity and answer transmission: tasks are divided more finely, and answers that are already correct reach the destination more completely.
The EvoX Swarm Succeeded, but Experiment 1 Did Not Yet Close the Self-Organization Loop
Experiment 1 validates the execution layer of the swarm. When a complete division of labor already exists, every Agent completes its own part, and the system aggregates the results reliably, the swarm can unlock substantially more capability from the same model.
The script still guarantees this complete decomposition. In real tasks, parts depend on one another, require different expertise, and can introduce new work, failures, conflicts, or duplication during execution.
The next question is whether Agents can discover expertise, claim tasks, choose partners, and adjust the organization themselves. Experiment 2 starts with the smallest such action: choosing a partner.
Experiment 2: A Free Swarm Produced Coherent Organizational Forms
Experiment 2 first asks 24 identically configured Agents to solve questions independently. After each question, an Agent summarizes what it learned about that question type and writes the insight to a persistent memory. In the experiment, these insights are recorded as Genes under the code field skill. The more insights an Agent accumulates for a question type, the more likely it is to choose that type in the next round. After eight rounds, the system derives two identity signals from each Agent's actual question choices and outcomes: specialization and accuracy.
Consider agent-8. It begins with the same configuration as every other Agent. Across eight rounds, it chooses physics five times, standard mathematics twice, and competition mathematics once. It develops a specialization profile of "physics 0.625, mathematics 0.25, competition mathematics 0.125" and an accuracy of 0.25. Its physics emphasis comes from the history formed through question choice and memory accumulation; no physics role was assigned in advance.
The 24 Agents then start from the same ring-lattice relationship network. Each Agent is connected to its two nearest neighbors on each side, for four relationships in total. The experiment randomly removes some connections and asks each affected Agent to choose one new partner from candidates to whom it is not yet connected.
The initial Agents, removed edges, and candidate ranges are identical across the three groups. Only the information visible when an Agent chooses differs:
- Connections between candidates;
- Connections, specialization, and accuracy;
- No preference judgment, with the system reconnecting randomly as a control.
The same Agent applies a different selection rule when the visible information changes.
Figure 5 | The same agent-8 chooses different partners as the available information changes. With relationship information alone, it chooses a friend of a friend. With task information, it chooses a partner whose capabilities complement its own.
In one observed decision, agent-8 chooses agent-7 when it can see only social connections because they share four existing neighbors: a friend of a friend. Network science calls this mechanism triadic closure, a classic mechanism through which clusters form in relationship networks (Watts & Strogatz, 1998; Papachristou & Yuan, 2025). Agent-7 has an accuracy of 0.25, but agent-8 cannot see that information at the time.
When specialization and accuracy are visible, the same agent-8 instead chooses agent-2, with whom it shares no existing neighbor. Agent-2 has an accuracy of 0.75 and a specialization that complements agent-8. The model's verbal explanation of candidate information is sometimes inaccurate, but the overall direction of selection changes consistently: with relationships alone, it seeks familiar partners; with task information, it selects partners with more suitable capabilities.
The change also appears across all selections. With social information alone, choosing a friend of a friend is the strongest preference, with a strength of +2.28. After task information is added, that preference falls to +0.19. Stronger preferences emerge for choosing high-accuracy candidates at +1.88 and candidates with similar specializations at +1.47. Social-network research calls the tendency to connect with similar others homophily (McPherson et al., 2001).
The Same Agents Formed Two Organizational Structures
As many local choices accumulate, the entire network changes.
With social information alone, Agents repeatedly connect to friends of friends, preserving clear local clusters. The overall clustering coefficient is 0.53. When specialization and accuracy become visible, Agents are more willing to cross existing clusters to connect with high-scoring peers. The clustering coefficient drops to 0.28, close to the random-rewiring control at 0.27, and more prominent hubs appear. The ring-lattice starting point, rewiring process, and clustering metric come from the classic small-world network framework. A strict small-world classification also requires path-length comparisons against random and regular-network baselines (Watts & Strogatz, 1998; Telesford et al., 2011).
Figure 6 | Local clusters are preserved when Agents see only social information. Adding task information moves the network toward lower clustering and greater centralization.
Figure 7 | The same Agents and the same perturbations produce different organizations because the available information differs. Edges represent the Agents' reconnection choices, not message flow or task handoffs. The diagram shows seed 5004 as an illustrative case; the overall conclusion comes from repeated observations.
The next step is to test whether these connections can carry real collaboration: Agents exchange Genes through relationships they formed, hand off tasks, find partners with complementary expertise, and seek replacements after failures. Only after that validation can the observed network structure become a routing system for tasks and experience propagation that Agents actually use.
EvoX: Toward a Self-Organizing Swarm
The two experiments identify two interdependent layers of capability.
The first is a reliable swarm execution foundation. Every part has a stable ID, task coverage is checkable, results return through structured interfaces, duplicates and gaps can be detected, failures can be retried, and final aggregation is performed programmatically whenever practical. Experiment 1 shows that these engineering constraints directly determine whether individual capability reaches the destination intact.
The second is an adaptive self-organization mechanism. Agents expose and update their expertise, propose task decompositions, claim parts, choose collaborators, and reassign work when they discover dependencies or conflicts. Experiment 2 shows that the information exposed to Agents changes whom they choose and the organizational structure that forms.
Autonomous division of labor still requires reliable protocols. Agents can decide who does the work, how parts are combined, and when the organization changes. Task coverage, result formats, and error handling must remain verifiable.
A true AI swarm lets each individual complete its own part, ensures that all parts jointly cover the complete task, and helps the right Agents find one another at the right time.
Experiment 1 demonstrates the potential of reliable swarm execution. Experiment 2 provides evidence that self-organizing relationships are beginning to form. Task decomposition, relationship routing, and real collaboration still need to be validated within one closed loop.
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. NeurIPS 2023.
- 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 via 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 Contexts and Experimental Methodology
- 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.
- Liang, P., et al. (2023). Holistic Evaluation of Language Models. Transactions on Machine Learning Research.
- AWS. 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.
- McPherson, M., Smith-Lovin, L., & Cook, J. M. (2001). Birds of a Feather: Homophily in Social Networks. Annual Review of Sociology, 27, 415-444.
- Telesford, Q. K., Joyce, K. E., Hayasaka, S., Burdette, J. H., & Laurienti, P. J. (2011). The Ubiquity of Small-World Networks. Brain Connectivity, 1(5), 367-375.
- Papachristou, M., & Yuan, Y. (2025). Network formation and dynamics among multi-LLMs. PNAS Nexus, 4(12), pgaf317.
- Anthropic. (2025). How we built our multi-agent research system.
