Agent Swarm

Agent Swarm is a "horizontal scaling" architecture that coordinates up to 300 sub-agents working in parallel — no predefined roles or hand-crafted workflows required. It completes tasks approximately 4.5× faster than single-agent execution.

On January 27, 2026, Moonshot AI released Kimi K2.5, introducing Agent Swarm. On April 20, 2026, Moonshot AI released and open-sourced Kimi K2.6, bringing major upgrades to the Agent Swarm architecture:

Up to 300 sub-agents working simultaneously
Over 4,000 tool calls per task
4.5× faster than single-agent sequential execution

Agent Swarm is currently powered by Kimi K3 (K3 Swarm), further improving large-scale parallel search and batch processing.

The story behind

In 2025, the AI industry's dominant narrative focused on vertical scaling — bigger models, more parameters. But this hits a structural ceiling: the single sequential execution bottleneck.

Agent Swarm was born from a real scenario: when a team member tried to automate daily stock information collection and hit 100 lines of if-else code, she realized: "I'm hand-writing a multi-agent system." If models can use tools, why can't they self-architect?

Agent Swarm is a self-designed organizational structure — designed by AI, not humans. The main Agent (orchestrator) autonomously directs up to 300 sub-agents, executing up to 4,000 parallel workflow steps.

Agent Swarm uses the PARL (Parallel-Agent Reinforcement Learning) training method. Compared to single-agent approaches, it reduces critical steps by 3×–4.5× in large-scale search scenarios.

How to use?

Access Points:

Web: kimi.com/agent-swarm
Mobile: open the Kimi app and select the K3 Swarm model in the model switch button

About the [Beta] tag: [Beta] indicates a trial marker for a product that is initially open to only a small number of users for small-scale validation and refinement; as the feature matures, the tag is gradually removed and the product is rolled out to everyone.

Agent Swarm is available to Moderato, Allegretto, Allegro, and Vivace members. Tasks consume significantly more credits than standard Agent tasks.

Steps:

Describe your task and send it (e.g., "Collect 200+ Paul Graham articles")
Watch real-time progress: task list creation, sub-agent spawning, parallel execution
Receive deliverables: code projects, file folders, data analysis, Office documents
Preview, download, or share results
In subsequent turns, Kimi automatically schedules between chat and Agent based on the task — no manual switching required

Use cases

Discovery at scale

Case 1: Top 3 Creators Across 100 YouTube Niches

Agent Swarm created 300 sub-agents for parallel search, generating structured tables with channel names, subscriber counts, and descriptions.

View result

Case 2: Collecting 200+ Paul Graham Articles

Agent Swarm deployed sub-agents to search, download, categorize, and summarize 200+ articles into thematic folders.

View result

Output at scale

Case: 100-Page Literature Review from 40 PDFs

Agent Swarm deployed multiple writing-focused sub-agents, each responsible for a chapter. Final output: a 100-page academic document with citations, methodology charts, and citation network analysis.

Perspective at scale

Case: Product Launch Strategy Expert Review Agent Swarm deployed expert sub-agents with different perspectives (Product Manager, Investor, Customer Success) to review a launch strategy.

View result

Case: The Three-Body Problem Rewritten in 20 Literary Styles 20 "writer" sub-agents composed independently in distinct styles — from Virginia Woolf to Borges to Kafka.

View result

Technical deep dive

Core Architecture: Commander + Specialists

Orchestrator = Coach/Commander: Sees the big picture, sets strategy
Sub-agents = Players: Each focused on a specific role

Key Design: Freeze the Players, Train Only the Coach

All sub-agents retain existing capabilities; only the orchestrator improves through reinforcement learning. This provides clear accountability and training stability.

Preventing "Laziness":

Serial collapse: Orchestrator handing everything to one sub-agent
Fake parallelism: Meaningless sub-tasks to game metrics

Solution: Three-Dimensional Reward Mechanism

Quality of final result
True parallelism achieved
Sub-task completion rate

Critical Steps Metric

Agent Swarm calculates the slowest sub-agent's time at each stage. This forces genuine process optimization rather than blind task splitting.

Context Sharding

Each sub-agent focuses on its own "notebook," recording relevant details independently. Only key conclusions are reported to the orchestrator — preserving reasoning without overwhelming memory.

Real-World Results

On the BrowseComp benchmark:

Accuracy: 15.9% (single agent) → 33.3%
Critical steps reduced by ~40%

Application scenarios

Agent Swarm is especially suited for:

Large-scale information retrieval: Massive internet data collection
Batch downloads: Large-scale file and resource collection
Wide-scope reading: Processing 100+ documents
Long-form writing: Content exceeding 100,000 words
Complex programming: Frontend development, code review, refactoring
Office automation: Professional documents, spreadsheets, presentations

Further Reading: