Moonshot AI's Kimi Work Unleashes 300 Agents to Run Tasks for Days

Kimi Work, Moonshot AI's local desktop agent, just got a significant upgrade: Goal Mode. The feature lets you hand the agent a high-level objective and walk away -- it keeps running, round after round, until the task is finished or you tell it to stop. It's the clearest signal yet that Moonshot is building toward something closer to an autonomous digital coworker than a chat assistant.

What Goal Mode actually does

Goal Mode lets you hand Kimi a goal and have it keep working toward that goal on its own. Instead of answering a single turn, Kimi saves the goal as a persistent state and, at the end of each round, decides what to do next: whether the goal is complete, blocked, paused, or ready to continue into another round. As long as your computer stays on, a goal can keep running across many rounds without you starting over each time.

This is a meaningful shift from how most AI tools work. Standard agent sessions are stateless -- you prompt, it responds, the context resets. Goal Mode breaks that pattern by treating the objective as a persistent first-class entity with its own lifecycle. The agent tracks whether it's making progress, what's blocking it, and what the next logical step is -- all without you having to re-engage after every step.

Rather than returning a one-shot answer, Kimi Work operates as a workflow loop: it assigns and executes sub-tasks across its agent swarm, verifies the results, and feeds them back into the next round of work. The entire process remains observable and open to human intervention.

The engine underneath

Kimi Work runs on Kimi K2.6, a roughly 1-trillion-parameter mixture-of-experts model with about 32 billion active parameters per token, a 256K-token context window, and an Agent Swarm system that scales to 300 sub-agents across up to 4,000 coordinated steps. Mixture-of-Experts (MoE) is an architecture where only a fraction of the model's parameters activate for any given input -- it lets you get the capability of a massive model without paying the full compute cost at inference time.

The 300-agent ceiling is a threefold increase from the 100-agent ceiling of its predecessor, Kimi K2.5. K2.6 can manage workflows spanning roughly 4,000 steps, up from 1,500 steps with K2.5. That's the difference between an AI that can draft a memo and one that can research a topic across multiple sources, cross-reference data, build a spreadsheet, write a summary, and schedule a follow-up meeting.

Moonshot says the model is designed for continuous execution, with internal use cases including agents that ran for hours and, in one case, five straight days, handling monitoring and incident response autonomously.

How to use it

Goal Mode is triggered with the /goal command inside Kimi Work. The official docs recommend a structured prompt format that includes your desired result, what counts as done (acceptance criteria), scope constraints, and what the agent should do if it gets stuck. Here's the template Moonshot provides:

I want to accomplish a goal: [your goal].
Please use /goal mode to keep working on it until you produce an executable, deliverable result.
Requirements:
1. Start by breaking the goal down for me.
2. Proactively ask me for any information you need.
3. Move forward step by step.
4. At the end of each round, give me the current progress and the next step.
5. Deliver a final, ready-to-use result.
My background/constraints: [add your situation]
Acceptance criteria: [what counts as done]

Goal Mode works best when these conditions are met:

• The goal is clearly defined but the path to get there is uncertain
• The task requires multiple rounds of attempts and verification
• The result can be checked against some concrete evidence (a passing test suite, a finished deck, a populated spreadsheet)
• The task would otherwise require you to re-prompt the agent after every step

A short report you can finish in one reply does not need Goal Mode. A research synthesis, a codebase migration, or a competitive landscape analysis does.

The full Kimi Work stack

Goal Mode is the headline, but it sits on top of a broader set of capabilities that make it useful. Kimi Work is a downloadable application, not a web chat. You give it goals in plain language, and it acts on your machine. The four core building blocks:

• Agent Swarm: Kimi Work can run many sub-agents in parallel on your machine. The swarm scales to 300 sub-agents. The system splits a task into parts, then coordinates the results.
• WebBridge: This browser extension lets the agent use a browser like a person. It searches, scrolls, extracts data, and fills forms across tabs. Because it uses your real session, it inherits your existing logins and cookies.
• Cron scheduler: A built-in scheduler runs jobs on a daily, hourly, or conditional basis. Triggers include LLM agent calls and Python or shell scripts. A "Keep Computer Awake" toggle keeps overnight jobs from stalling.
• Local file access: The agent reads folders you mount and runs Python in the background. Original files stay in place unless you approve a change.

The app also ships with native market data for A-shares, Hong Kong stocks, and U.S. equities pre-integrated -- no API setup required.

Where it shines (and where it doesn't)

The strongest use cases are tasks that mix local data, live web sources, and multi-format output -- things that would normally require you to juggle several tools manually:

• Finance and investment: Pull earnings reports from a local folder, cross-reference live price data, flag concentration risks, and generate a morning briefing deck -- automatically, every day
• Academic research: Fan out 300 agents across hundreds of papers simultaneously, extract methodology and findings from each, and synthesize a structured literature review draft
• Codebase migrations: Set the acceptance criteria as a clean build and passing tests, then let Goal Mode iterate until it gets there
• Competitive analysis: Profile dozens of companies in one run, pulling live signals from public pages and combining them with offline documents
• Scheduled reporting: Pull weekly campaign metrics every Friday, compare against historical data in a local spreadsheet, and deliver a finished review deck before Monday

The limitations are real, though. When your laptop is closed, tasks stop. Moonshot's own cloud product, Kimi Claw, runs 24/7 without your machine. Because WebBridge drives your actual logged-in browser, it can touch your bank account, your email, and your company's internal tools. Researchers at UC Riverside warned in May that AI agents often carry out tasks without recognizing when their actions are risky -- a behavior they called "blind goal-directedness." The "ask before acting" mode is the right default to leave on.

Availability and pricing

Kimi Work is a local desktop agent for macOS (Apple silicon) and Windows. The app is a free download. Meaningful agent features require a paid plan.

Moonshot's Moderato tier starts at $19/month and includes K2.6, Deep Research, and Kimi Code access. The Agent Swarm with limited sub-agents unlocks at Allegretto ($39/month), while the full 300-agent swarm and highest-volume professional workflows require the Allegro ($99/month) or Vivace ($199/month) tiers.

The bigger picture

The desktop agent race has gotten crowded fast. Anthropic's Claude has offered full-desktop computer use since late 2024. OpenAI shipped Codex Background Computer Use for macOS in April 2026, running agents in parallel desktop sessions. Google's Gemini computer use -- descended from Project Mariner -- focuses on browser workflows. Microsoft's Copilot Studio added computer-use in May 2026, aimed at enterprise automation.

The difference with Kimi Work is the local-first design paired with a 300-agent swarm. Most competitors either run everything in a cloud sandbox -- which can't touch real logged-in sessions -- or offer desktop control without coordinated parallel agents. Kimi Work does both. The tradeoff is machine dependency, but for workflows that live in local files and authenticated browser sessions, that's often an acceptable constraint.

Kimi Work is powered by Kimi K2.6, a natively multimodal model that Cursor -- one of the most popular AI coding tools in the Western market -- quietly adopted for its new coding features in March 2026. That kind of third-party adoption is a stronger signal than any benchmark. Goal Mode is the feature that turns a capable local agent into something that can genuinely own a task end to end -- and that's the bar the whole industry is now racing toward.