Worked continuously for up to 40 days. The Silicon Valley Agent has evolved again: Give it an instruction and it'll handle the rest.

An all - automatic intelligent agent that can independently plan and execute continuously for 40 days is here!

Sure enough, when it comes to eating lobster meat, sometimes we still have to look at foreigners (doge).

The newly launched Missions by Factory directly surpasses OpenClaw, presenting a plate of shelled meat on the table -

No more empty talk! Just give a single task instruction, and it can deliver a full - fledged automatic engineering closed - loop.

In other words, whether it's a complex architecture or cross - module tasks, give it a Missions, and the intelligent agent can plan, write, and test on its own.

Just focus on the results, not the process. Users no longer need to constantly stare at the screen for interactive instructions.

To be honest, it's really awesome!

Can work continuously for up to 40 days

According to the official description, Missions is installed on Factory's self - developed intelligent agent Droids and is designed to autonomously manage complex tasks spanning multiple days.

Users only need to tell the intelligent agent what they want to do, such as "Help me build a CRM system" or "Migrate this PHP codebase to TypeScript". The intelligent agent can automatically break down the task into subtasks and execute them in chronological and logical order.

Each subtask will generate a corresponding dialogue, coordinate and hand over through Git, and verify and fix errors in a timely manner at each step, finally producing a complete result directly.

Meanwhile, in the terminal, you can see the task progress of the intelligent agent in real - time, including the functions being built, which intelligent agent is executing, and which tools are being used.

Specifically, to enable multiple intelligent agents to run in parallel, Missions has a built - in scheduler.

The scheduler breaks down large - scale complex projects into multiple milestones. In each milestone, the work is further subdivided into multiple functions.

Each function will start a brand - new context dialogue window to avoid losing context or making errors in a single long - running dialogue. At the same time, under appropriate circumstances, Missions will perform parallel processing within the functions to improve efficiency.

The prerequisite for moving on to the next milestone is that the previous milestone must complete the verification phase.

The system will review the completed work, run tests, and verify whether all functions have been integrated. Once problems are found during the verification process, the scheduler will automatically generate repair tasks until the standards are met, and then proceed to the next stage.

Missions has built - in native computer functions and is specifically optimized for task workloads. The verification process will be aligned with human verification.

In addition, Missions itself supports multiple models and can call AI models from different manufacturers and of different types to act as execution intelligent agents. The scheduler Droids will automatically select the most suitable model according to the task.

In addition to software development, Missions can also be used for tasks such as training machine - learning models and writing research papers, with strong generalization ability.

This is mainly achieved through a skill - based learning system. When a new task runs, it will extract reusable operations into skills. Then the execution intelligent agent will continuously improve and expand this skill library while working.

The more users use it, the better the system will perform in the user's field.

With the addition of Missions, the duration of the intelligent agent's dialogue has increased significantly.

Previously, the average dialogue duration of Droids was about 8 minutes, and 60% of the dialogues were completed within 15 minutes. However, the dialogue between Missions + Droids generally lasts about 2 hours, and 37% of the tasks last more than 4 hours.

Some tasks even last for several days, and the longest task can reach 40 days.

This means that the introduction of Missions enables the intelligent agent to handle more complex tasks, greatly raising the upper limit of AI autonomy, truly at the top level in the industry.

Meanwhile, not only does the task runtime become longer, but the number of inferences per round also increases.

In a single Mission, the message sending rate drops to 3 messages per minute, but the token weight of each message doubles. This is because Missions spends a lot of time on executing engineering tasks rather than constantly generating text tokens.

On average, Missions consumes 12 times the number of tokens as a normal dialogue, but the consumption speed before and after is actually about the same, about 45,000 tokens per minute.

Currently, Missions is available in the official CLI and IDE extensions. Users of the enterprise version and Max version can use it starting today.

The powerful combination of theoretical physics and AI

Factory.ai is a Silicon Valley startup. Different from traditional AI code assistants (such as GitHub Copilot), it is committed to building autonomous AI engineers.

Its representative product is Droids, an autonomous agent designed specifically for the software development lifecycle.

It can independently complete complex tasks, understand user needs, consult documents, write code, and submit, covering all aspects of software development.

Similar products, such as the previously popular AI programmer Devin, but Droids emphasizes deeper integration into the enterprise workflow.

This time, Missions is a system - level encapsulation and scheduling upgrade for Droids, establishing a complete multi - agent coordination framework.

The founders of the team are two Princeton University alumni, Matan Grinberg and Eno Reyes.

Matan Grinberg is a typical example of transitioning from a theoretical physics background to AI entrepreneurship. During his undergraduate studies, he studied under the world - top string theory expert Juan Maldacena. When pursuing his Ph.D. at UC Berkeley, he studied the intersection of physics and AI.

In 2023, with an email full of in - depth thoughts on string theory and AI, he caught the attention of Sequoia Capital partner Shaun Maguire. So he dropped out of school to found Factory and received millions of dollars in investment led by Sequoia.

Eno Reyes has a rich background in machine learning. Before founding Factory, he worked as a machine - learning engineer at Hugging Face, responsible for model - related work, and also engaged in software development at Microsoft.

He majored in cognitive science at Princeton University and is responsible for leading the research and development of Droids' autonomous cycle mechanism and context compression mechanism.

The two of them reunited at a hackathon and decided to start a business. They signed a letter of intent in one day, registered the company, and the initial demo only took a short week.

Forbes once evaluated this combination as the golden partnership in the AI era, and Shaun Maguire also praised them:

They not only have top - notch physical and AI capabilities but also have matching business intuition and execution efficiency.

Their concept is to further stimulate creativity in the field of software development while AI is developing strongly - not just limited to engineers.

Reference links:

[1]https://x.com/FactoryAI/status/2027104794289263104?s=20

[2]https://factory.ai/news/missions

[3]https://sequoiacap.com/article/partnering-with-factory-autonomous-ai-for-all/

This article is from the WeChat official account "QbitAI", author: Luyu. Republished by 36Kr with permission.