Kimi launches Agent Swarm: up to 100 subagents and 1,500 tool calls

Moonshot AI has added an Agent Swarm mode to Kimi K2.5. Given a task, the model decides on its own how to split it, spawns up to 100 subagents, runs them in parallel and spends more than 1500 tool calls on the work. According to the company, the result arrives 4.5 times faster than with sequential execution. The feature is released as an early research preview and is currently available only to top-tier subscribers.

The ceiling of a single agent

The starting diagnosis in the Kimi blog is that the main constraint today is not the model's intelligence but the sequential way a single agent works. Ask a deep research tool to survey a hundred companies or pull together dozens of articles, and the context window gradually fills up. The system starts compressing and summarizing its history to free up room for new tokens. That compression is lossy, and the later reasoning steps get worse. Moonshot calls this a structural ceiling, not a temporary bug.

The company's answer is to scale not only up (more parameters, higher benchmark scores) but also out. A single agent works like one expert, while a network of self-organizing agents is closer to a small company.

How it came about

The backstory is modest. In June 2025, shortly after the launch of Kimi Researcher, a team member who is into trading decided to automate her daily collection of stock data. She needed to check news on macro trends, pull the history of "limit-up" counts and dig deeper when the conditions matched. She wrote a cascade of if-else statements in Python, and around line 100 realized she was building a multi-agent system by hand. That raised a question: if a model can use tools and hold long tasks together, why can't it decide for itself when to parallelize work, whom to "hire" and how to delegate?

In Agent Swarm, the main agent acts as a manager. It identifies the workers it needs (researchers, analysts, fact-checkers) and creates them for the specific task. The user does not have to define the orchestration.

Specs

• Model: Kimi K2.5
• Parallel subagents: up to 100
• Tool calls per task: more than 1500
• Speed: 4.5 times faster than sequential execution (company data, no independent measurements)
• Status: early research preview, available to top-tier subscribers
• Planned: direct communication between subagents and dynamic control over the width of parallelism

What Kimi shows

The examples in the blog fall into three types. The first is search at scale. The agent identifies 100 niche YouTube topics and launches a separate subagent for each to find the top 3 creators. In another example, the swarm collected more than 200 Paul Graham essays scattered across personal sites, old blogs and transcripts of talks, and sorted them into six thematic folders with a final report.

The second type is large-volume output. 40 PDFs on social psychology became a 100-page, two-column literature review with formatted references. Author subagents take on individual sections, and the texts are then merged.

The third type is the most interesting: deliberate disagreement. A product plan is evaluated by a skeptical venture investor, an experienced PM, an ethics specialist and a customer success lead. In a literary experiment, 20 "writers" in styles ranging from Virginia Woolf to Borges and Kafka continue Liu Cixin's The Three-Body Problem. The blog's authors claim that independent agents reach different conclusions, which then have to be reconciled. In their words, this structurally protects against groupthink.

How to try it

A swarm makes sense where the work genuinely parallelizes: broad research, batch uploads, processing many files, multi-perspective analysis, long texts. If the steps depend tightly on each other, the gain will, judging by the logic of the approach, be noticeably smaller. That is our conclusion, not a Moonshot claim.

There are also practical caveats. More than 1500 tool calls mean significant token consumption, and the blog itself half-jokingly measures AI-era literacy by the number of tokens spent. It is worth spot-checking the results: a hundred agents can be confidently wrong in parallel. Finally, subagents do not yet talk to each other directly, which is one of the announced improvements. A good starting point is a task with a clearly structured output: a table, a set of folders, a report with sections. That makes it easier to judge whether the claimed speed multiplier holds up.