Agent Swarm

A multi-agent setup where one orchestrator model spins up many specialized sub-agents that work in parallel on parts of a large task (as described for Moonshot Kimi K2.6).

How it works

An agent swarm is architected around a central orchestrator LLM that receives a complex task, decomposes it into subtasks, and spawns specialised sub-agents to handle each subtask in parallel. Sub-agents have access to tools (web search, code execution, file access) and report results back to the orchestrator, which synthesises them into a final output. The orchestrator also manages inter-agent communication and handles failures by re-delegating or retrying tasks.

Why it matters

Agent swarms are the current frontier of agentic AI capability. They allow AI systems to tackle tasks that are too large, complex, or multi-domain for a single agent in a single context window. Moonshot AI reported using swarms of 2,048 parallel sub-agents in Kimi K2.6 for complex coding and research tasks. As context windows remain finite and specialised agents outperform generalist ones on narrow tasks, swarm architectures represent a practical path toward AGI-level task completion.