2026-08-29 · Agentic Realism

The tool changes. The desk stays.

Why the value is no longer in the latest assistant

The model is a capability. The harness is a commodity. The method — framing the work, guaranteeing the result — is the craft. For a Swiss SME, the tool can change. The thread, the gates and the last word cannot.

Overhead editorial illustration: a stable desk, a notebook, a compass, a leaf and copper threads; around it, a pencil and a USB stick — tools move, the post stays.
Editorial illustration for this article. The desk is the place that remains; the tools around it can change. It is not a product interface or proof of an installation.

I. One month, then another tool

A landscaper in Vevey learned, in February, to dictate quotes into an assistant. In April a colleague said “everyone has moved on”. In June a course promised the real secret: a new coding tool, faster, more agentic. In the evening the same slope photos were waiting. The same follow-ups. The same lost thread between the call, the quote and the work.

This is not a story of incompetence. It is a story of layers. He was sold a tool as if it were a craft. He was sold a capability as if it were a method.

Agentic Realism, Response-able AI starts elsewhere. A useful agent is not the latest software you learned. It is framed work, in a place that remains, with a human who can still say no.

II. Three layers, only one that holds

In work with agents, three things sit on top of each other. We mix them too often, because the market likes to sell the layer that changes fastest.

The model is a capability. It reads, it proposes, it drafts. The following week another one leads on a benchmark. That is useful. It is not a practice.

The harness is the loop that runs the agent: permissions, tools, sub-agents, takeovers. In a few months it has become a commodity. Like an IDE. You can switch. You do not build a firm on it.

The method is what organises the work after the model and the harness have already changed. Aims. Limits. Artifacts. Gates. Review. Last word. If it is not written down, it dies with the tab.

  1. 1ModelA capability. It leads, then another overtakes it.
  2. 2HarnessThe loop: tools, permissions, sub-agents. An IDE, not a craft.
  3. 3MethodFrame the work. Guarantee the result. Outlive the next tool.
Diagram: three layers. The model and the harness change. The method, if written down, remains.

The model is a capability. The harness is a commodity. The method is the craft.

We already knew this for drawing software or bookkeeping. We forget it as soon as the assistant speaks. Speech gives the illusion of a craft. It is not one.

III. Frame, then guarantee

A therapist in Lausanne wants an agent to prepare reminders, never the content of a session. If she starts with “build me a system”, the agent invents a world that is too wide: notes, mail, memories, files. If she starts by framing, the world is small and true: slots, invoices, already-approved messages. Reading is not drafting. Drafting is not sending.

Two skills hold from one year to the next.

Framing is designing the work before releasing it: aim, perimeter, artifacts, what stays out of frame. A quote. A site photo. Not the family album.

Guaranteeing is controlling the result: roles, gates, feedback, takeover. Who may read. Who may prepare. Who may commit a send, a payment, a signature.

  1. 1FrameAim, limits, artifacts — before the agent runs.
  2. 2ExecuteThe agent prepares, inside the perimeter.
  3. 3GuaranteeGates, review, human last word.
Diagram: two durable skills. Designing is not executing. Executing is not committing.

A prompt is not a method. A method is not a piece of software. And software is not an agential cut — the boundary that decides what enters the agent’s world, and what does not.

IV. What we refuse

We refuse to train a practice on a tool that will be stale in six months. We refuse the title that promises a magic replacement, then reveals, at the end of the video, that there was no tool: only method. The trap was the title. The market lives on that trap.

We also refuse the inverse: despising tools. A good harness helps. A fairer model helps. They remain lower layers. We test them. We change them. We do not put our memory in them.

Above all we refuse the hustle that says: clone any service in a day, launch twenty agents without gates, spend three subscriptions so as not to “miss the train”. In French-speaking Switzerland a small firm does not need a train. It needs a thread.

If everything lives in an assistant, everything dies with the assistant. What must last gets written down: files, rules, logs, decisions. That is already the sense of intra-action: intelligence is not in an isolated model. It emerges in the apparatus — human, files, limits, cadences.

V. A desk that outlives tools

An artisan who changes assistant every quarter does not have a model problem. He has a place problem. Photos, quotes, follow-ups must remain somewhere when the software changes. A useful agent needs a desk — not only a capability.

The desk is not a harness. It is the post: files, limits, screen, human takeover. You can bring the agent you already use. You do not have to rebuild your method with every fashion. The tool changes. The desk stays.

AlpenDock, when it serves, serves that: a computer for the agent, isolated, observable, with a human who takes the hand back. Not a religion of tool. A post. Data residence is then stated honestly, workspace by workspace — never as a slogan.

Conclusion: keep the thread

Train for framing and for guaranteeing — not for the latest tool. Bring the agent. Hold the desk. Let the model and the harness do their work as lower layers. For a therapist, an artisan, a solopreneur, the value is not knowing how to launch this month’s software. It is designing the work, controlling the result, and still being able to answer for what goes out.

The tool changes. The desk stays. Framing, gates, last word: yours.