You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$50,000 left to reach the funding goal.
Automaton is an open-source operating system that gives each agent its own seat ensuring agents and people can use the same computer without compromising for the other. It is being designed with agents' and humans' needs in mind.
Multiple problems we have with agentic workflows today stem from the fact that the agent and the human user of the computer are the same entity from the operating system's point of view. The traditional approach to solving this problem has been to sandbox the agents. However, this has the flaw that one has to manage multiple applications, and the average person outside of the software ecosystem probably won't bother setting all of this up.
All these can be solved by redesigning the operating system to account for the fact that there are other users of the computer that are not humans, and placing all of the security systems in place from the operating system level, rather than by applications. Automaton is meant to do just this. The approach is to give identities to agents, and separate access to computer resources using the different identities. Letting the user choose what their agents are allowed to read, write, or modify, and knowing that this is all enforced at the fundamental operating system level means people can let agents work on their computer without having to trust them.
I built a working demo on Linux myself. In it, six agents ran six long computer-use tasks at the same time on one PC with 8 GB of memory and no graphics card, while a person kept working beside them, and nothing the agents changed reached the original files until it was reviewed and merged. Building it showed me a few problems that would be cleaner to solve deeper in the operating system. For these three months I'll keep building on Linux, and I'll write those problems up as work for later.
Demo video: https://youtu.be/Ugkh9cJ_mVg
Here are the goals I have for the project, and a timeline for execution.
Give agents user identities of their own
Separate agents from human users
Let human users choose the scope of access their agents get
Over three months:
Month 1: one installable Linux-based image with the agent seat built in.
Month 2: a test agent actively tries to break the isolation, reading files it wasn't given or reaching outside its seat, and I fix what it finds.
Month 3: other people install it, and I fix what breaks for them.
All security features meant to separate the agent access from human access work during tests.
Several agents execute tasks that cannot be done on today's operating systems at once, example have multiple agents do computer use at the same time.
At least ten people besides me install it and use it for a week.
The money pays for two things: the compute to build and test Automaton, and a stipend for my time so I can work on it full-time instead of taking a job.
At the $30,000 minimum (3 months, full-time):
$15,000: stipend for my time, so I can work on it full-time for three months
$8,000: a rented GPU running a smaller open-source model as a test agent that actively tries to break the isolation, reading files it wasn't given or reaching outside its seat
$4,000: API tokens for the agents doing normal computer-use tasks
$3,000: two or three cheap test laptops with different hardware, so it doesn't only work on my machine
At the $50,000 goal:
$25,000: stipend for my time, so I can work on Automaton full-time for five months
$16,000: a larger GPU for the full three months, so the test agent is a stronger model and a harder test
$6,000: API tokens
$3,000: test laptops
I'm Ronald Kouatchoua Djakou, a Mathematics sophomore at the California Institute of Technology, and I'm working on Automaton alone. My body of work has primarily been computational research in polymer chemistry, quantum physics, and mathematics research on cellular automata.
The most likely causes:
Ordinary apps behave badly when they run on an agent's separate desktop, so the seat only works for some programs.
Agents stay too slow or unreliable at computer use for people to bother running several at once.
A large OS maker ships the same idea first.
Even if Automaton itself doesn't take off, the pieces are still useful on their own: the agent isolation, the review-and-merge step for agent changes, and the results of the test agent trying to break out. All of it will be public.
There are no bids on this project.