Swarm Intelligence: Understanding Kimi Agent Swarms
Swarm Intelligence: Understanding Kimi’s Multi-Agent SystemHow one orchestrator turns hundreds of AI agents into a temporary society for research and discovery.Photo by and machines on UnsplashMulti-agent systems usually use a small team of specialized agents to solve complex, long-horizon tasks.…
Swarm Intelligence: Understanding Kimi’s Multi-Agent SystemHow one orchestrator turns hundreds of AI agents into a temporary society for research and discovery.Photo by and machines on UnsplashMulti-agent systems usually use a small team of specialized agents to solve complex, long-horizon tasks. Moonshot AI’s Kimi K3 Agent Swarm operates at a different scale: it can spawn more than 300 agents and analyze a catalogue of gravitational waves.1. Gravitational Waves and Agent SwarmsOne of Moonshot’s demonstration cases for K3 Agent Swarm is a scientific analysis of GWTC-5, the fifth Gravitational-Wave Transient Catalog. This cosmological catalog contains signals from mergers of black holes and neutron stars detected by the LIGO, Virgo, and KAGRA observatories.To understand what one entry represents, imagine two black holes or neutron stars spiraling into each other. As they orbit, they generate gravitational waves, ripples in spacetime that can be detected on Earth by gravitational wave detectors.Analyzing one event means comparing that signal against template signals to estimate the source’s physical parameters: the masses, the spins, the distance, and the sky coordinates.GWTC-5 contains 391 independent events, and the scientifically interesting questions are about the statistical population of all events, how black-hole masses are distributed, whether their spins align, and how the black-hole merger rate changes with cosmic distance from Earth.Kimi K3 GWTC-5 gravitational-wave analysis (presentation preview)Kimi K3 used around 20 sub-agents to analyze GWTC-5 events, produced seven scientific visualizations, two tables, and synthesized findings from 10+ research papers.For more examples of explicit role-based multi-agent systems used in scientific discovery, see the related article on Automated Discovery with Multi-Agent Teams.2. What Is Swarm Intelligence?In agent swarms, the focus shifts from explainable individual agents to the behavior of organizations of agents.The term swarm intelligence predates language models. It emerged from research on collective behavior in ants, bees, birds, fish, and distributed robotic systems.In a classical swarm model, many individuals interact locally, without a central controller specifying every action. Useful macroscopic group behavior emerges from these microscopic interactions.For example, in an ant colony, individual ants respond to local signals such as pheromone trails, food, congestion, or the behavior of nearby ants. No individual has a full model of the colony, yet the colony can allocate labor, find optimal routes, and even solve complex physical and logistical problems.Artistic illustration by Aron Visuals on UnsplashA classical swarm is defined by decentralized control, local interactions, and self-organized behavior. This is characterized by:many interacting agentslocal information or decisionsdecentralized coordinationself-organizationnew collective behavior that emerges without a predefined plan.Kimi’s swarm architecture does not perfectly fit the scientific ‘swarm’ definition. It has a central orchestrator that interprets the task objective, creates sub-agents, delegates sub-tasks, and integrates results. This can still be called an agent swarm in a broader sense, but more precisely it is an orchestrated swarm.In contrast, natural bee swarms do not have a central coordinating bee. Coordination emerges from local interactions among many bees. The queen reproduces, but she does not assign tasks or control collective decisions.Many modern papers now discuss centralized, hierarchical, and hybrid swarm architectures. These are also called swarms, but they are not the classical models inspired by ants and bees.3. Parallel-Agent Reinforcement LearningK2 Swarm was introduced by Moonshot AI along with a training method called Parallel-Agent Reinforcement Learning (PARL). This training method is used to teach the orchestrator how to organize work across agents.During training, the orchestrator learns when to create sub-agents, how to decompose a task, and how to schedule subtasks in parallel. The orchestrator is also an agent — specifically a trainable coordinator — not a script. It receives the user’s objective, creates and assigns sub-agents, monitors their progress, and combines outputs into a final result.In multi-agent orchestration, there are two common mistakes. The first is serial collapse. The orchestrator can learn that creating and coordinating other agents is difficult, so it defaults to doing everything itself. It technically possesses swarm tools but rarely uses them.The second is spurious parallelism. The orchestrator spawns many agents because the training process appears to reward parallel activity, even when the additional work is redundant or meaningless.PARL is designed to avoid both of these failures. During model training, the sub-agents are frozen, and only the orchestrator is updated. This separates the ability to perform a subtask and the ability to organize many subtasks.The orchestrator agent receives a main performance reward for the quality of the final result. It also receives temporary auxiliary rewards that encourage exploring sub-agent creation and completing tasks. These incentives are gradually reduced so that the final policy optimizes the actual outcome rather than the superficial appearance of collaboration.4. Large-Scale Agent SwarmsRecent research on multi-agent scaling found strong diminishing returns when many homogeneous agents produce correlated outputs.Diversity in models, prompts, tools, or evidence matters more than population size. Two genuinely complementary agents contribute more than twenty agents with the same reasoning.Large agent swarms are less helpful when every step depends on the previous one. This is the same limitation from ordinary parallel computing.5. Benchmarks and WorkspacesOn the BrowseComp benchmark, which requires deep web research and synthesis, the reported score increased from 60.6 for a single K2.5 agent to 78.4 for an agent swarm.On WideSearch, which measures broad information seeking, item-level F1 increased from 72.7 to 79.0. On Moonshot’s internal Swarm Bench, covering wide retrieval, more than 100 input documents, and very long-form writing, performance increased from 41.6 to 58.3.In WideSearch experiments, Swarm reportedly reached target performance 4.5 times faster than sequential execution. The strongest performance occurred on benchmarks designed for wide parallelism.Moonshot also introduced Claw Groups, shared workspaces where multiple persistent AI agents, called Claws, collaborate on a goal. Claw workspaces are different from agent swarms; they coordinate already configured agents with persistent capabilities.6. Society of AgentsHuman organizations exceed individual capability not because every member becomes smarter, but because work, memory, expertise, criticism, and verification are distributed across many people.One agent may identify a promising scientific hypothesis, another may search the literature for supporting evidence, while a third compares conflicting results. As new evidence appears, all agents revise their conclusions.This resembles the behavior of scientific teams and research laboratories. Intelligence may scale not only by building a stronger model, but by teaching a model to create and manage a temporary society of agents.TL;DR: Kimi’s Agent Swarm is a centrally orchestrated multi-agent system. It is trained via PARL. K3 supplies the newest long-context model and 300+ agents. The swarm improves decomposable tasks by increasing parallel coverage and distributing context, but agent count alone does not increase intelligence.ReferencesKimi Team, “Kimi K2.5: Visual Agentic Intelligence” (2026). https://arxiv.org/abs/2602.02276Moonshot AI, “Kimi K2.6: Advancing Open-Source Coding” (2026). https://www.kimi.com/blog/kimi-k2-6Moonshot AI, “Kimi K3: Open Frontier Intelligence” (2026). https://www.kimi.com/blog/kimi-k3E. Bonabeau, M. Dorigo, and G. Theraulaz, Swarm Intelligence: From Natural to Artificial Systems (Oxford University Press, 1999).S. Garnier, J. Gautrais, and G. Theraulaz, “The Biological Principles of Swarm Intelligence,” Swarm Intelligence 1, 3–31 (2007).Y. Yang et al., “Understanding Agent Scaling in LLM-Based Multi-Agent Systems via Diversity” (2026). https://arxiv.org/abs/2602.03794M. Cemri et al., “Why Do Multi-Agent LLM Systems Fail?” (2025). https://arxiv.org/abs/2503.13657This story is published on Generative AI. Connect with us on LinkedIn and follow Zeniteq to stay in the loop with the latest AI stories.Subscribe to our newsletter and YouTube channel to stay updated with the latest news and updates on generative AI. Let’s shape the future of AI together!Swarm Intelligence: Understanding Kimi Agent Swarms was originally published in Generative AI on Medium, where people are continuing the conversation by highlighting and responding to this story.Source: Generative AI Pub — Published — Category: Image AI