Kimi, the AI company behind the popular Kimi assistant, has unveiled Agent Swarm, a new system that reimagines AI collaboration. Instead of relying on a single agent, Agent Swarm can deploy up to 100 specialized sub-agents that work in parallel, autonomously organizing themselves like a company with bosses, employees, and division of labor.
According to Kimi, the system is designed to overcome the limitations of single-agent reasoning, particularly for long-horizon tasks. As tasks grow, context windows fill and reasoning degrades—a structural ceiling that Agent Swarm bypasses by distributing work across multiple agents.
The idea originated from a Kimi team member who tried to automate daily stock research using a Python script. After writing over 100 lines of if-else logic, she realized she was essentially hand-coding a multi-agent system. This sparked the hypothesis: if models can use tools and handle long tasks, why can't they architect themselves?
Agent Swarm is now available as an early research preview for top-tier subscribers. It excels in tasks that benefit from parallelization, such as broad research, batch downloads, multi-file processing, multi-angle analysis, and long-form writing. Kimi reports that Agent Swarm delivers results 4.5x faster than sequential execution.
Example use cases include finding the top 3 creators in 100 niche YouTube domains, collecting all 200+ Paul Graham essays scattered across the web, generating a 100-page literature review from 40 PDFs, or having a team of expert personas review a product launch plan from different perspectives.
Kimi emphasizes that Agent Swarm is not just about speed—it creates conditions for productive disagreement, avoiding groupthink by having independent agents arrive at different conclusions and then reconcile them. The company plans to continue hardening the architecture with direct sub-agent communication and dynamic control of parallel width.