There is a term catching on right now. The self-driving company. A business where AI agents do not just assist with the work but carry it, end to end, across every function. Research runs itself. Marketing runs itself. The product gets built, shipped, and supported while founders continue to iterate on what's next.
We have been building toward this for a while, so we want to say clearly what we think it means, what it does not mean, and where the hard problems actually are.
The levels, borrowed from driving
Self-driving cars gave us a useful vocabulary. Nobody argues about whether a car is autonomous. They ask what level.
Level 0 is manual everything. You do the work, the tools hold still.
Level 1 is assistance. Autocomplete, grammar checks, a chatbot that answers questions when asked. The human does the driving, the software smooths the ride.
Level 2 is the supervised copilot. The AI produces real work, but a person watches every step and approves every move. This is where today's copilot tools are designed to sit. It feels like autonomy. It is actually supervision with better output.
Level 3 is conditional autonomy. An agent takes a goal and runs it end to end. It plans, executes across dozens of steps, and comes back with finished work. The human is on call, not in the loop of every action, unless they want to be.
Level 4 is domain autonomy. Whole functions run themselves inside explicit boundaries. Research keeps itself current. Marketing ships on schedule. Ops watches itself. People set direction and handle what the system escalates.
Level 5 is full autonomy, no humans anywhere. AGI is achieved. We believe the technology gets there eventually. We also believe nobody is there yet, and anyone selling it today is selling ahead of reality.
Most of what is sold as AI autonomy today is level 1 or 2 by design. The self-driving company conversation is really a conversation about leaning into levels 3 and 4.
Capability is not the blocker
Here is the uncomfortable part. We think frontier models are already capable enough for level 3 work across most business functions. The blocker is not intelligence. It is trust.
An agent operating autonomously reads web pages, opens documents, and consumes API responses. Any of that content can carry adversarial instructions. Research published in 2026 found that intent-hiding attacks succeed 85 to 100 percent of the time against tested systems (arXiv 2603.04474). And in multi-agent workflows the damage compounds. Five of six tested frameworks reached total system failure from a single injected error.
A company will not hand a function to an agent it has to babysit. Supervision is exactly what levels 3 and 4 are supposed to remove. So the whole concept lives or dies on one question. Can the agent be trusted to stay inside its lane without a human watching?
Self-driving needs brakes the driver did not build
Our answer is the same one the automotive world reached. You do not put the safety system inside the thing being controlled. A car's braking system does not ask the engine for permission to work.
Mode Agent enforces boundaries at the orchestration layer, outside any model's inference. Every action an agent proposes passes through enforcement before it executes. What it can access, what it can do, who it can delegate to, how long it can run. No web page, document, or tool output the model reads can change that, because enforcement happens in a separate system that reads none of it.
That is what makes the higher levels real instead of a demo. Autonomy you can verify, not autonomy you have to hope about.
The human keeps the wheel where it matters
A self-driving car still needs someone to say where to go. Nobody wants a taxi that picks its own destination.
The same is true of a self-driving company. Direction, taste, and accountability stay human. You decide what the business is, what it will not do, and what good looks like. The Agent executes the whole playbook underneath that. It validates the idea, maps the competitors, writes the positioning, runs the launch, builds the product, and knows what comes next without being asked. You define the boundaries. It does the work inside them.
That is not us hedging on what AI will become. It is how a self-driving company works at today's level of the technology. Autonomy without human direction, at the current state of the art, is a car with no passenger, driving nowhere in particular, very efficiently.
Where we actually are
Honest scorecard. Mode Agent runs level 3 today. Give it a goal and it carries whole workstreams end to end, in parallel, across research, marketing, engineering, and more, inside enforced boundaries. We are building toward level 4, where functions stay continuously current on their own, and we are measuring the enforcement layer as we go rather than asserting it works. That research is ongoing and we will publish what we find.
Level 5 is an AGI question. We believe AGI is possible, and we are building so that when the technology arrives, the boundaries and the verification are already in place. We are not there yet. Neither is anyone else.
For how the Agent runs work end to end, go to gotmode.com/how-it-works. For the enforcement architecture, go to gotmode.com/safety. For the research behind it, go to gotmode.com/r-and-d.