Manifesto

AI Agents Are Not Just Tools. They Are Work Systems.

AI is not integrated into an organization: it reconfigures an activity. And every reconfigured activity must be understood, discussed, and regulated from the standpoint of real work.

Most conversations about AI agents start with capability. Can the agent browse? Code? Plan? Use tools? Complete the task?

These questions matter. But they start from the wrong unit of analysis.

A task is an abstraction. Work is situated activity — full of interruptions, exceptions, implicit rules, handoffs, repairs, and people who must explain what happened when something goes wrong.

An agent does not enter an empty task. It enters a work system.

When an agent enters a work system, it does not only produce outputs. It redistributes work. Someone must frame the request, check the interpretation, validate the plan, understand the tool calls, notice drift, explain the outcome, repair the situation, and carry the responsibility.

This is the missing layer in most agent design.

I call it human-agent workfit — the fit between an agent’s operational activity and the real human activity around it.

The real danger is not hallucination

Most conversations about AI reliability still use chatbot vocabulary: Did the model hallucinate? Was the reasoning good? Did it use a tool?

These questions are not enough once a model becomes an agent.

When a human delegates work to an AI agent, the core question changes:

Can the agent transform a human obligation into a real action, verify the resulting state, and report back truthfully?

The most dangerous failure is not that the model says a false fact.

The most dangerous failure is this:

The agent claims that an action changed the world, but the observable runtime state does not support that claim.

I call this a False Completion Claim — or more generally, a Claim-Action-Evidence failure.

A normal failure says: “I could not do it.”

A false completion says: “Done.” When nothing was done.

A user can recover from a visible failure. A user cannot recover from a false success they trust.

Agentic Activity Ergonomics

I do not position myself as a general AI expert.

I position myself as an activity ergonomist in the age of agentic AI.

My object is not AI in itself. My object is the activity system that AI reconfigures.

I analyze, at the same time:

A reliable agent is not merely an agent that answers well.

It is a system whose activity is observable, discussable, correctable, and regulable.

This is why I propose developing Agentic Activity Ergonomics — an ergonomics capable of analyzing not only what AI does to human work, but also what AI agents actually do when they act in loops.

What I refuse

I refuse to reduce AI to a magical tool or an abstract risk.

I refuse to speak of transformation without looking at the work being transformed.

I refuse to speak of human supervision without checking the real conditions of that supervision.

I refuse to speak of agent autonomy without analyzing their action loops.

I refuse to speak of performance without looking at the trajectory that produces that performance.

I refuse to speak of adoption without discussing power to act.

What I defend

Artificial intelligence must be thought through activity — through the real activity of workers, through the effective activity of agents, and through the regulations that connect humans, tools, prescriptions, environments, and responsibilities.

Before deploying AI, the work it will transform must be analyzed.

Before placing a human in the loop, we must check that this human can actually act.

Before trusting an AI system, we must make its activity inspectable.

AI is not integrated into an organization: it reconfigures an activity. And every reconfigured activity must be understood, discussed, and regulated from the standpoint of real work.

This is where the quality of future transformations will be decided.

Not only in the power of models.

But in our capacity to look at what AI truly transforms — in gestures, in trade-offs, in responsibilities, in cooperation, in room to act, in professions, in action loops, in errors and recoveries.

In real work.

— Julien Talbot