The environment defines the agent

Aug 31, 2026

I wrote these words in reaction to several people in the last few days asking me what I think about Grok Bot and the current hype around it.

I don't think "AI employees" are a particularly useful abstraction for agents.

A lot of agent products right now basically ask you to recreate a company inside the software. You make a "Steve", tell Steve that he works in marketing, give him a personality, a job description and some tools, and then Steve talks to your sales agent and your research agent and eventually you have a little org chart of bots coordinating with each other.

And tbf it makes intuitive sense because companies are made of, duh, people. Agents can increasingly do things people do and so the obvious thing is to model agents as people. But I think this we might be getting this wrong.

We keep rebuilding the old environment inside the new technology

New technologies often start by imitating whatever they replace.

The first cars looked like carriages without horses. Early websites looked like magazines. We still call things on computers "files", put them into "folders", throw them into a "trash can", and for a surprisingly long time the easiest way to explain a smartphone was basically a phone that could also run applications.

This isn't stupid. Familiar metaphors are incredibly useful when nobody understands the new thing yet.

However, I think something similar is happening with agents/LLMs.

We know how human organizations coordinate intelligent actors today: employees. Employees have names, roles, responsibilities, managers and org charts. So when LLMs becomes capable of doing increasingly large pieces of work, the immediate instinct is to put it into the same structure.

You create an AI SDR, an AI researcher, an AI marketing manager. Maybe they get names, profile pictures and some quirky way of talking. You define what each one does, give them access to some tools and eventually let them talk to each other.

The company stays basically the same. We just start replacing some of the boxes on the org chart with LLMs + system prompt + tools.

But if agents actually become capable over long time horizons and large context, why would the organizational primitive designed around humans remain the right one for software?

What actually defines an employee?

Think about what makes someone at a company useful.

Imagine hiring a new marketing person.

You can give them a job description describing exactly what their role is, but that description contains almost none of the information they actually need to do the job.

They need to know what the company is building. What customers have been saying. What was tried 6 months ago and didn't work. What the CEO cares about. Which claims legal doesn't want you to make. What engineering is shipping this quarter. What the current campaign looks like. Who owns design. Where the latest assets are. Which conversations they should be part of. What they are allowed to publish without asking somebody.

Basically their actual "definition" is spread across the environment. Aka Slack channels, meetings, documents, permissions, tools, relationships, previous decisions and thousands of tiny pieces of accumulated context. And over time the person becomes better at their job because they become more embedded in that environment.

This is why the idea that we should separately define every agent starts to feel strange to me.

If all of the information that determines how an agent should behave already exists in the environment, why are we recreating it in the agent? Or better, why do we even have to explicitly create and maintain an agent?

The environment can define the agent

Imagine instead that the computer is basically the company.

There is a filesystem containing the conversations, artifacts and history of the company. People exist inside it. Permissions exist inside it. Tools exist inside it. Work happens inside it.

Maybe there is a marketing directory/space/folder (whatever you want to call it). That space contains the people working on marketing, every relevant conversation, the current website, previous campaigns, customer research, analytics, brand assets, permissions to publish somewhere, tools to generate things and whatever else marketing accumulates over time.

Now intelligence acts inside that space, coordinates and gets the job done ideally.

What exactly do we gain by first creating a fictional employee called "Steve", writing a system prompt explaining that Steve is a marketing employee that talks like Alex Hormozi, giving it access to Buffer or whatever and then asking Steve to work?

The whole environment already tells the model what it is. It screams it actually.

An AI operating inside /marketing should behave like it belongs to marketing because it sees the world from there. The same underlying intelligence operating inside /finance should behave differently because the context, people, history, permissions and available actions are different.

The context is the definition!

And I think this gets more interesting as agents become more capable, because you can stop thinking of the agent as a persistent little person that travels around carrying its identity with it. You can actually think of intelligence as something that gets instantiated against a particular environment when work needs to happen.

This changes coordination too

The AI employee metaphor doesn't only give agents identities. It imports a lot of the coordination structure of human organizations with it.

If Steve is the marketing agent and Sarah is the research agent, Steve has to ask Sarah to research something. Sarah returns the research to Steve. Steve might send something to the design agent, which sends something back, and eventually the agents need some protocol for figuring out who should do what and when.

There is quite obviously real work happening here, and multi-agent systems can be useful. But sometimes I look at these architectures and wonder whether we are accidentally rebuilding the coordination overhead we spent the last 100 years creating because humans have extremely specific limitations.

Humans cannot instantly inherit arbitrary context. We have bounded memory and we specialize because learning everything takes years. We need meetings because information is distributed between brains. We need managers partly because somebody has to maintain shared state across those brains. We create departments because 5,000 people cannot all maintain relationships with each other.

A lot (maybe all, not sure) of organizational structure is just downstream of those constraints.

Agents have constraints too, obviously. But they aren't necessarily the same constraints :)

If the same underlying intelligence can enter another context and immediately read its history, artifacts, permissions and current state, why does it need to "talk to" another fictional employee to get there?

Maybe sometimes it does, but that should probably be a consequence of the work, not the fundamental architecture.

This is where I think the distinction between agents and environments becomes important.

Agents probably become more ephemeral

There are obviously things that should be persistent. Like, a process that monitors production should continue running, a sales process might own a set of accounts over months. Some agents will need memory, explicit instructions, identity, budgets, credentials or a very specific set of capabilities.

And maybe we call all of those things agents. The terminology isn't that important imo. But I suspect the average unit of AI work becomes much more ephemeral than the "employee" metaphor suggests.

Something happens -> Intelligence gets invoked inside the relevant context -> It has the state it needs because the state belongs to the environment -> It does the work, changes the environment and disappears.

The durable thing in this case isn't necessarily the agent, but the environment and the state inside it.

That also makes the system much easier to reason about. Instead of asking what Steve knows, what Sarah knows, whether Steve told Sarah something, whether Sarah's memory is stale and which version of the company each agent currently has in its context, there is one shared reality that humans and AI operate against.

This seems much closer to how computers already work.

We don't create a little database employee with a personality and ask it to remember our customer records. We create a database. Computation acts against it.

Maybe agents end up looking less like digital people and more like computation acting against increasingly rich organizational state.

The company itself probably has to change

This is the part I personally find most interesting because it connects to a broader pattern with new technology.

For example, humanoid robots make a lot of sense today because the physical world is built for humans. Doors, stairs, kitchens, tools, cars and warehouses all assume roughly human geometry. A humanoid robot is basically backwards compatible with the existing world.

But over a sufficiently long time horizon, if robots become economically important enough, we probably won't keep every environment exactly the same and endlessly optimize robots to imitate humans inside it. Factories already don't work that way. We change the environment around the machine. And once you can change both sides of the equation, very different forms become possible.

I think we're currently doing something similar with AI inside companies. The company is built for humans, so we make the AI look like another human participant. Give it a name, give it a role, put it into Slack, give it an inbox, let it attend meetings, let it message other agents etc. etc. etc.

This is probably necessary during the transition because all of the work is already organized that way. But eventually the more interesting question is what the company itself looks like once you assume both humans and capable machine intelligence from the get go.

And I don't think the answer is a normal company with 300 fictional employees added to the org chart. Maybe explicit AI employees are a useful transitional metaphor and maybe some persist permanently for things where identity itself matters.

But I suspect that as the underlying systems get better, we will care less about designing the character of each agent and much more about designing the environment in which intelligence acts.

The agent, in this case, might increasingly just be a function of the context around it.