Kimi launches 100-agent parallel swarm system

The structural limits of a single agent

As AI reasoning systems try to handle longer tasks, single-agent models inevitably hit a wall. If a single agent is tasked with researching hundreds of companies or synthesizing dozens of papers, its context window fills up as the work progresses, and the system summarizes and compresses earlier records to free up space. This compression process is lossy, 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 imposed by context windows, 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 design their own organization

Agent Swarm is not simply a matter 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 this as "an organizational structure of bosses, staff, and departments—designed not by humans, but by the system itself."

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 was also designed to structurally prevent collective bias, by having independently operating agents reach different conclusions and then reconcile them.

Three use-case scenarios

Moonshot AI presented three cases where Agent Swarm proves especially effective: large-scale discovery, large-scale output, and multi-perspective analysis.

As an example of large-scale discovery, in a task to find top creators across 100 YouTube niche domains, the K2.5 Agent Swarm autonomously generates 100 sub-agents per domain to conduct 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, organizing them into six topic folders. For large-scale output, an example was given of feeding in 40 social psychology PDFs to generate a single 100-page academic document complete with footnotes and references. For multi-perspective analysis use cases, a complex product launch plan can be reviewed by a team of expert agents—including a skeptical venture capitalist, a veteran PM, an ethicist, and a customer success representative—each from their own perspective, or 20 writer agents with distinct literary styles can each continue writing an ending to a novel.

Availability and future plans

Agent Swarm is currently available as an early research preview to users on Kimi's top subscription tier. The announcement did not disclose specific subscription pricing or tier criteria.

While Moonshot AI stated that the current architecture is ready to handle real tasks, it also noted that this is not the finished version. Direct communication between sub-agents and dynamic control of parallel width are planned for future addition. No restrictions on supported regions or languages were disclosed separately.