이미지: 문샷 AI
Summary
- Kimi Agent Swarm is a multi-agent system that deploys up to 100 sub-agents in parallel and handles more than 1,500 tool calls.
- Its core design principle is overcoming a single agent's context window limits and sequential execution bottlenecks through horizontal scaling.
- It is currently available as an early research preview to users on the top subscription tier, with direct sub-agent-to-sub-agent communication and dynamic control of parallel width planned for future addition.
The structural limits of a single agent
As AI reasoning systems attempt to handle longer tasks, single-agent models hit a consistent wall. If a single agent is tasked with researching hundreds of companies or synthesizing dozens of papers, the context window fills up as the work progresses, and the system summarizes and compresses previous records to free up space. This compression process involves loss, and the quality of subsequent reasoning degrades as a result.
Moonshot AI framed this not as a bug or a temporary flaw but as a structural ceiling created by context window size, time, and reliance on a single agent. The starting point for developing Agent Swarm was the recognition that vertical scaling alone — faster inference or lower cost — cannot break through this ceiling.
Horizontal scaling: agents designing their own organization
Agent Swarm is not simply a method of running multiple AI agents together. When a user requests a task, the system organizes its own sub-agents, distributes roles, and executes them in parallel. Moonshot AI described it as "an organizational structure of bosses, employees, and departments designed by the system itself, not by humans."
In numbers, the Kimi K2.5-based Agent Swarm deploys up to 100 sub-agents simultaneously, executes more than 1,500 tool calls, and delivers results 4.5 times faster than sequential execution. The structure also aims to structurally prevent collective bias by having independently operating agents reach different conclusions before coordinating them.
Three use-case scenarios
Moonshot AI presented mass discovery, mass production, and multi-perspective analysis as scenarios where Agent Swarm is particularly effective.
As an example of mass discovery, in a task to find top creators across 100 YouTube niche domains, K2.5 Agent Swarm autonomously generates 100 sub-agents for each domain to perform parallel searches. Another example cited was collecting, categorizing, and summarizing more than 200 of Paul Graham's essays from his personal site, blog posts, and talk transcripts, then organizing them into six thematic folders. For mass production, an example was given of feeding in 40 social psychology PDFs to generate a 100-page academic-style document, including footnotes and references. For multi-perspective analysis, the system can have a team of experts — a skeptical venture capitalist, a veteran PM, an ethicist, and a customer success manager — each review a complex product launch plan from their own perspective, or have 20 writer agents with distinct literary styles each continue writing the ending of a novel.
Access conditions and future plans
Agent Swarm is currently offered as an early research preview to users on Kimi's top subscription tier. Specific subscription pricing or tier criteria were not disclosed in the announcement.
Moonshot AI stated that the current architecture is ready to handle tasks, while making clear it is not yet a finished version. Direct communication between sub-agents and dynamic control of parallel width are planned as future additions. No restrictions on supported regions or languages were disclosed separately.
