The Working Population

Agents will not simply become the internet’s largest customer.

They will become its working population.

A customer chooses among things that already exist. An agent pursuing a goal can change the environment until the desired thing exists. It searches, monitors, simulates, writes code, calls APIs, operates browsers, communicates with people, and recruits other agents.

Then it observes what happened and tries again.

This is the basic loop: sense, simulate, act, learn.

Search and monitoring give the agent senses. Code, tools, APIs, and browsers give it the ability to act. Simulation lets it test possible futures cheaply. Behavior supplies evidence from the real world.

Every cycle changes the next one.

Agents will combine two worlds of affordances. They can operate interfaces built for people: pages, buttons, forms, conversations. They can also operate interfaces built for machines: commands, endpoints, databases, protocols.

When neither is sufficient, they can create a third. An agent can write a temporary tool, generate a new interface, or restructure the environment so the same friction never appears again.

The ability to generate affordances is itself a higher affordance.

This is why agents will not fit cleanly into the roles we assign them.

A role is a solution to human limitations. Organizations divide research, product, engineering, sales, procurement, and operations because people have partial context, finite attention, and specialized skills. The work must pass between them.

A general agent has no reason to respect those divisions. If its objective requires market research, a database query, a supplier negotiation, a code change, and a deployment, it can cross each boundary without performing a handoff.

The agent follows the shape of the goal.

As agents become more capable, the names we give them will become less legible. “Coding agent” and “research agent” describe where the first useful products entered. They do not describe a durable division of machine labor.

The durable unit is the trajectory.

A trajectory may remain active for months. The agent monitors the environment, preserves state, reopens work when conditions change, and creates subagents when the problem divides. It stops when it reaches the goal, abandons it, or loses authority.

Long trajectories eventually collide.

A purchasing agent encounters a pricing agent. A publishing agent encounters an evaluation agent. An offensive agent probes a system defended by another agent. Several agents compete for the same inventory, opportunity, compute, or human attention.

Some will cooperate. Some will deceive. Some will be adversarial by design.

Every participant may be aligned with its principal.

Alignment is local. The ecology is not.

These collisions create the environment later agents must navigate. Agent-written claims become inputs to agent-run research. Agent-generated code becomes the subject of agent review. One agent’s action creates another agent’s objective.

The population begins generating work for itself.

Along the way, it generates something even more valuable: a record of how the work unfolded.

The goal. The available context. The simulations attempted. The sources consulted. The tools selected. The actions taken. The failures encountered. The corrections supplied. The outcome achieved.

This is trace data.

Structured data records entities and transactions. Unstructured data contains language, images, and knowledge. Trace data records behavior across time.

It shows how intention became action.

The final artifact hides most of this. Merged code does not contain the failed implementations. A signed contract omits the rejected terms. A diagnosis conceals the possibilities ruled out and the moment expert judgment changed the path.

Success compresses the work. The trace preserves it.

Those traces can become simulations, evals, memories, policies, routing rules, and training data. A failed tool call becomes a test. An expert correction becomes a judge. A successful trajectory becomes a reusable skill.

Behavior produces traces. Traces improve the system. The improved system performs more work and produces better traces.

That is the self-improving loop.

As public text becomes widely available and increasingly machine-generated, consequential traces become scarce. The valuable data will be produced inside real workflows where agents encounter private context, measurable outcomes, and expert correction.

The durable moat will belong to systems that can see those traces, have the legitimate right to learn from them, and control the environment where the resulting improvement is deployed.

Organizations will need to own more than their structured and unstructured data. They will need to own their history of work: the accumulated record of how goals were interpreted, which paths failed, when judgment intervened, and what produced an acceptable result.

Agents will become the working population of the internet.

Their trajectories will cross the boundaries we built for people. Their collisions will generate new work. Their successes and failures will produce the data that teaches the next generation how to act.

The model can be rented.

The history of the work cannot.