0
Maximum parallel agents
0
Maximum coordinated steps
0x faster
Than single agent on complex tasks
Step 1: Task Decomposition
You provide a complex task and Kimi breaks it into independent sub-tasks
Step 2: Agent Spawning
Up to 100 parallel sub-agents are launched, each assigned to a different aspect of the problem
Step 3: Parallel Execution
Agents work simultaneously, searching the web, analyzing data, and generating findings
Step 4: Coordination
Agents share intermediate results across up to 1,500 coordinated steps using learned strategies (PARL)
Step 5: Synthesis
All findings are combined into a unified, comprehensive response
| Agent Swarm | Single Agent | |
|---|---|---|
| BrowseComp | 78.4 | Lower |
| WideSearch | 79.0 | Lower |
| Speed on complex tasks | ~7 minutes | ~30 minutes |
| Coordination | Learned via PARL | Sequential only |
What makes Agent Swarm different from competitors
Most competitor agent systems use a single agent working sequentially or a small number of agents with predefined roles. Agent Swarm's scale (100 agents, 1,500 steps) and its use of learned coordination through PARL rather than hand-crafted workflows are genuinely novel. The parallel architecture means complex research tasks that would take a single agent 30 minutes might complete in under 7 minutes.
Beta limitations
Agent Swarm is still in beta. Results can be inconsistent, and it works best for research and analysis tasks. Creative or subjective work is less reliable. The coordination overhead means it is not beneficial for simple tasks. Use Instant or Thinking mode for straightforward questions.