Moonshot's Kimi K2.5 runs 100 parallel sub-agents

Moonshot AI introduced Agent Swarm, a mode for its Kimi K2.5 model. Kimi K2.5 acts as an orchestrator: it breaks a request into parts, creates specialized sub-agents on the fly and merges what they return.

According to the company, Agent Swarm can:

• run up to 100 sub-agents in parallel
• make more than 1,500 tool calls
• deliver "better results 4.5x faster than sequential execution"

The post gives no benchmark, baseline or method behind the 4.5x figure, and it has not been checked independently.

Moonshot says the problem it is solving is structural. When a single agent works through a long research task, its context window fills up. It then has to fold or summarize its history, which loses information, and its later reasoning gets worse. Running many agents side by side spreads that load across separate contexts. The company also says independent agents that reach different conclusions help avoid "groupthink."

The examples Moonshot gives include:

• finding the top three creators in each of 100 niche YouTube categories, with one sub-agent per category
• collecting more than 200 Paul Graham essays and sorting them into six topic folders
• turning 40 social psychology PDFs into a 100-page literature review with formatted citations
• reviewing a product plan from several assigned viewpoints, such as a skeptical VC and an ethicist

Moonshot traces the idea to a team member who was writing a hand-coded Python workflow for gathering stock information. About 100 lines in, she realized she was "hand-coding a multi-agent system."

The post links to shared sample sessions. The material reviewed does not cover pricing, rate limits or how failures between sub-agents are handled.