Agent Swarm
An agent swarm is a coordinated group of AI agents running simultaneously, each handling a sub-task, with results combined by an orchestrating system into a single output.
A single AI agent handles one task at a time: read this file, write this function, call this API. An agent swarm runs many agents in parallel, each assigned a different piece of a larger problem, coordinated by an orchestrator that assembles their outputs. Moonshot AI’s Kimi K2.6, released today, can orchestrate up to 300 simultaneous sub-agents on a single coding task.
The practical difference between one agent and a swarm is coverage and speed. A single agent working through a large codebase refactor might take hours and make sequential decisions that compound errors. A swarm can analyze different modules simultaneously, surface conflicts in parallel, and return a synthesized result in a fraction of the time. The tradeoff is complexity: coordinating 300 agents requires careful task decomposition, and errors in the orchestration layer can multiply rather than cancel.
For operators building with AI, agent swarms matter because they shift the ceiling on what can be automated in a single session. Tasks that were too large or too multi-threaded for a single agent pass, like a full codebase migration, a multi-source research synthesis, or an end-to-end QA cycle, become tractable. The cost structure changes too: running 300 agents simultaneously burns tokens at a rate that requires careful budgeting. Understand the agent loop cost before running a swarm at production scale.