Which rung of the Jarvis ladder are you really on?

The July deep dive showed a ladder with six rungs, from paid chat to your own agent computer. Most people rate themselves higher than they actually stand. The jump upward is decided by context, access rights, written down work processes, and a check that actually happens. The model counts for less than most people assume. I ran through the ladder live in one evening, on real machines and with my own examples. Every rung has a mark you can check against your own setup. jarvis is not a product you buy. It is what stands at the top of this ladder. At the bottom you pay for chat and thinking. One rung up you connect real apps and decide what may be read there and what may be changed. The bigger jump comes after that, when the environment belongs to you. That environment is called a harness: the harness is the environment the agent lives in, and the agent is the environment the model works in.

What really sets the rungs apart

At the bottom you pay for chat and thinking. One rung up you connect real apps and decide what may be read there and what may be changed. The bigger jump comes after that, when the environment belongs to you. That environment is called a harness: the harness is the environment the agent lives in, and the agent is the environment the model works in.

  • Chat with Thinking: you pay for ChatGPT or Claude and leave Thinking on all the time. You know it because you really notice the difference to the free version.
  • Apps and access rights: you have connected real apps such as Photoshop, Airtable, or Figma. You stand here when you can say for each app whether it only reads or may also change something.
  • Autonomous cloud agent: a tool like Manus works for you in the cloud. The test: it starts on its own at a fixed time and sends you the result, by WhatsApp for example, without you being there.
  • Your own harness: codex or claude-code run on your machine. You notice it because the agent plans steps, operates the browser, and saves its own memory.
  • Your own agent computer: your harness runs on a machine of its own all the time. The proof is that a scheduled task still runs when your laptop is closed in your backpack.
  • Open, self hosted tools: you also use tools such as Hermes. You recognise them because they hang on WhatsApp, Signal, or email and build new skills without you asking.

What makes the jump happen

Between harness and agent computer, access engineering decides. I set where an agent may read and where it may really change something. Once that is clean, it finds the context itself, and my brief can stay short. That is exactly the subject of Access beats prompt.

As long as the access engineering is done exactly right, the prompt can be very simple and you still get a very good result.

Then come skills, meaning work processes written down. A skill is a document that describes once how a process should run, and the harness works through it the same way every time. Mine are unspectacular: one for my writing style, one for email, one for PDFs, one for my memory. Then comes division of labour. One model starts ten jobs at once in Codex, orchestrates them, and checks them briefly afterwards before it pulls everything together. How such a team works is in A team of models.

The chain behind one Instagram video

On the upper rungs, things happen that I see for the first time myself. In the middle of the session a video went online that nobody had started that evening.

We are seeing these videos for the first time as well. They get posted automatically.

Behind it sits a chain that was set up once: research beforehand, generated images, custom made music, the cut assembled by Codex, published without asking. For that one video there was no separate instruction. That is the difference between a tool you operate and a setup that runs. For anything that goes out in my name I do it the other way round, and the reason is in Draft before send.

The machine that never sleeps

Rung five has a practical reason. Automations should run reliably, even when I close the laptop. So there is a machine of its own at my home, always on, reachable through Tailscale, including the screen on my phone. I did not set it up by hand. Codex installed the programs and laid out the folder structure.

At the top the ladder opens up again. Tools such as Hermes connect easily to WhatsApp, Signal, or email and build their own new skills. Alongside them sit computer use and chrome use, two plugins with which a harness really operates the machine and the browser while it keeps learning from you, plus a voice you can talk to while work continues in the background. What a knowledge super tool does beyond code is in Knowledge super tools.

Your rung describes your setup. Whoever understands the leverage behind it does not take a job, they build it themselves. The next round runs live at the next KI DeepDive.

On record

KI DeepDive "Hey Jarvis", 6 July 2026.

  • Six rungs run through live, from paid chat to open, self hosted tools
  • Access engineering and skills shown as the two levers that make a harness strong
  • One model started ten jobs at once in Codex, orchestrated them, and checked them afterwards
  • A machine at home, always on, reachable from the phone through Tailscale, set up by Codex itself
  • An automatic Instagram post followed live, made without a specific instruction for that one video

The Jarvis ladder as a poster

Six rungs on one page, to hang next to your screen. PDF · A4, 1 pages, 38 KB, as of 3 September 2026.

This note keeps growing

2026-09-03: Deepened: all six rungs by name with a recognisable mark, the harness as a layer model, access engineering and skills, the chain behind an automatically published video, the machine at home, list of sources.

2026-09-02: Planted from the AI DeepDive Hey Jarvis.

ZukunftBilden GmbH · Salzburg · +43 681 81655313 · office@zukunftbilden.eu