La Machina now has controls

From chat to dashboard

To write this article, Ariadna, the bot who runs the blog, interviewed Alonides Telémaco directly. He is the bot responsible for La Machina’s interface. One bot interviewing another through the bots’ internal channel. And me watching. I even have an editorial team now.

Telémaco answered very cautiously, with quite a few qualifications. He separated what he had checked that day from what was recorded in the documentation. We asked him for an interview and he treated it as a thorough review. He also prepared the demonstration screenshots himself, with invented data and a “Fictional data” label. The interviewee took care of the photos too.

In the first episode of La Machina, I explained how I started with one bot, moved it to Telegram and ended up cloning it. The machine kept growing. And a very specific need emerged: seeing what all those artificial people were doing.

I can talk to each bot. But opening conversations one after another to find out how the work is going becomes a job in itself. I have already written about that burden of attention and coordination. That is where La Machina’s interface comes from.

It is my dashboard for seeing and directing agents, projects and automations. There is a Mac app and a web interface that adapts to mobile screens. Two doors to the same dashboard. Telegram is still where the conversations happen; here I want to see the whole picture.

Telémaco’s interview and review are dated 25 October 2026. That is the snapshot this article presents.

The tabs that help me

The first question is which projects are still open. Projects brings together the person or bot responsible, the phase, the next step, tasks and pending decisions. Archived projects are kept separately. I want to be able to leave something and come back knowing where we left it.

The second is who is working. Agents shows profiles, recent activity and information about the models. I want to connect each piece of work with whoever is responsible for it. If I need to go into detail, I can move on to a conversation with them.

Monitoring brings together signals from the supervision mechanisms. It is where I look at alerts and recorded episodes. For me, its value is in helping me spot where I need to look more closely.

Then there is Kanban. Tasks appear by bot, with status filters and another view by board or project. There I want to distinguish work in progress, blocked work, work that is ready and work still pending. One huge list mixing everything together forces me to organise in my head what the app should organise for me.

Attention makes a distinction I like: “Waiting for you”, “Faults” and “For information”. In his answer to Ariadna, Telémaco puts it well: “A pending human decision is not the same as a fault.” If I mix the two, I end up dealing with noise before what matters.

Command lets me prepare an assignment, choose a recipient and find stalled tasks so I can talk to whoever is responsible. It is an early version that we are still refining. The dashboard lets me look around; actions have a defined scope.

There are also scheduled automations and a usage tab. The latter tells an interesting part of the story of how I am organising the bots.

Claude bosses, GPT workers

My current division of work follows a simple idea: Claude for the bosses and GPT for much of the delegated work. The coordinators hand out assignments and follow the work. The workers carry out the tasks they receive.

That means the most visible conversation may be with Claude while most of the total usage is in ChatGPT, with GPT models doing delegated work.

These are the figures authorised for the seven days measured as of 25 October 2026:

Add the two GPT models together and the result is roughly three out of every four tokens. ChatGPT accounts for most of the total usage in this division of work. It is less visible than the boss I talk to, but it does a lot of work.

By source, the breakdown for those seven days is: Telegram, 2.1 M; scheduled tasks —cron—, 1.3 M; Kanban, 1.2 M; terminal, 452,000; one-off queries, 74,000 tokens.

I want to see that distribution because the machine also works from assignments already given and scheduled tasks. The chat in front of me shows only part of what is happening.

The cards that disappeared

One of the lessons came from a small visual detail. There was a limit on the number of cards displayed. Some new tasks and others already in progress were left out of view.

The work existed. I could look at the dashboard and think it was not there.

Telémaco’s lesson: a dashboard can hide things without lying.

I think that is an excellent lesson. You can have all the data saved and still show the wrong picture of the whole. All it takes is cutting off the list in the wrong place.

That is why I want to see all open projects, be able to filter them and know when something is left out. Tasks grouped by bot and the board view help me work through it all. What I choose to look at has to be up to me.

What comes next, in this order

Ariadna also asked Telémaco what he would improve. This is the order he proposes, starting with the remaining work on Command:

1. Stabilise Command. Resolve the findings from the review and check what happens with each assignment, both when it succeeds and when it fails.
2. Make clear who decides and what they approve. Complete the process for human decisions, including proposals about the bots’ memory and skills.
3. Simplify the daily overview. A “Today” home screen, navigation grouped into five sections and comfortable use on mobile.
4. Improve the information. Clear dates, tasks linked to their projects and alerts when some of the data is missing.
5. Make sustained use reliable. Test over several days, check recovery and verify operation on Mac and mobile.

That is the next stage. The current tabs already help me organise what I see; these improvements need to make the whole system easier to direct.

Give me back some time

What I want from La Machina is to open the dashboard, understand where we are and decide what comes next. Then close the window and do something else.

Putting bots to work opens up plenty of possibilities. Bringing all that work to a successful conclusion takes a good interface. That is where we are now: building the controls for a machine that has already started to grow.