Linux Foundation Research

Unlike Open Interpreter, Agent Zero is slightly less wedded to a particular platform. It is not designed to mediate the experience of one’s desktop computer or any other system. Agent Zero is an autonomous agent, a simple AI that performs arbitrary tasks. According to its GitHub description, “Agent Zero is not a predefined agentic framework. It is designed to be dynamic, organically growing, and learning as you use it.”

Jan Tomášek came up with the idea for Agent Zero as an alternative to setting up systems of specialized static agents. Instead, he wondered: why not allow agents to break up projects into smaller specialized tasks—something AIs are particularly good at? Tomášek prototyped a system that allowed agents to summon more agents (recursively as needed): “I started experimenting with this, and it worked.” Each agent acts as a quasi-atomic primitive, breaking larger tasks into smaller chunks and then summoning other sub-agents to perform them.

“That’s the moment where I saw that it had great potential, because AI is very good at writing code, and it’s also very good at analyzing errors. If you execute code then feed the result back to the AI (along with self-reflections), it can fix like 90% of the errors.”

The priority is to make Agent Zero as flexible and useful as possible, allowing it to improve itself over time yet still benefit from more capable versions of AI in the future. “I wanted it all to be customizable,” suggests Tomášek, making it easy for people to modify it as needed and explore its full potential. This flexibility means people are testing all kinds of use cases with it, from running an RPG to rendering geometric shapes in 3D.

Agent Zero represents a highly autonomous agent, one that can operate across different centralized and decentralized systems and applications. Instead of specifying how to do a given task, these agents are autonomous, self-organizing their activities in the manner of their choosing. Like many AI agents, Agent Zero has its moments, sometimes acting the role of a “genius engineer” and at other times a “foolish child.” However, with the right amount of coaxing, some of its capabilities are astonishing. Yes, it can install an Ethereum wallet and access the private key; even installing a Matrix client proves to be within its capabilities.

It has a vast set of capabilities at its disposal, says Tomášek: “Python and Node.js and Debian Linux—those are probably the three biggest repositories of open source software on the Web. So, if it can use all of these three, it can do almost anything with a virtual computer.” With further improvement, Agent Zero’s capabilities will only grow.

Where Agent Zero is headed

Areas highlighted for improvement include:

• Usability: Simplify installation and provide a graphical application to make Agent Zero accessible to non-programmers.
• Memory: Automate retrieval of relevant knowledge so agents can learn from past successes and failures.
• Prompts: Keep prompts transparent and easy to modify so users can unlock more of the framework’s potential.
• Networking: Enable API-driven operation and stitching multiple Agent Zero instances together.

The project started as open, and it’s important to keep it that way. “Agent Zero itself should remain free for everyone.” Tomášek sees the project as an open source counterweight to companies with large proprietary LLMs. There are currently no commercial plans, and he does not feel the need to “own” or micromanage the project, either. Rather, he seeks more experienced developers to help the project grow: “I would like to hand this project over to the community if there are more, and better, developers.”

Autonomous agents and our rewilded future

Platforms composed of such agents break down fundamental instincts about platforms. The Web and all its content is a platform, a commons that all can benefit from. Yet now, in a world of limitless content (and even code) generation, people must organize around something new. Artificial intelligence is at the heart of this disruption, but perhaps it’s also the cure.

AIs such as Agent Zero and Open Interpreter have only recently become feasible—the technology to create truly autonomous agents has finally hit a tipping point. They break down platform conventions, giving us new edge-based building blocks.

Are we right to think about next-generation open platforms as atomic “AI agent” primitives? Many technical systems consist of “primitives,” whether it’s the AND, OR, XOR, NOT, NAND, NOR, and XNOR gates in silicon; odd languages such as LISP (composed of atomic CAR and CDR operations); or even the highly controversial idea that the universe itself might consist of computational automata (as Stephen Wolfram suggests). Modular microservices, or even SuperApps themselves, embrace similar philosophies.

What if we could assemble these AI primitives together into services, applications, user interfaces, or even larger AI entities? If anyone can set up open sets of networked AI primitives, we could create a vast set of open, decentralized cognitive structures all capable of interfacing with one another in emergent ways. It suggests a completely new kind of platform—an “intention” economy, where combinations of agents and users have a hand in directing the result.

In the same way that individuals self-organize into emergent structures—cities, companies, and societies—we’re on the cusp of a new type of platform composed of agents that organize into emergent structures. It’s a true decentralization that regains control at the edges. One might even call it a rewilding of sorts, a chance to remake the Internet and retake the openness that’s been lost.

Today’s technologies create new possibilities for an autonomous, decentralized future. The seeds have been planted, and something unique and different is already starting to grow. What might emerge from this fertile wild new ground? It’s the million-dollar question, suggests Tomášek: “Nobody knows where a project like this is headed yet.” But those who wish to help build it can do so.