Knowledge pillar
Behind the scenes, episode 05: L'Alter, from answers to assistance
I used ChatGPT and Claude to get answers. Then curiosity pushed me to look behind the interface: context, memory, tools and permissions. I was not looking for another chat window. I needed an assistant for the concrete tasks of everyday life.
2026-10-25 — 7 min read
Listen to this article
Like many people, I started with the best-known assistants.
I asked ChatGPT or Claude a question. I got an answer, carried on the conversation, then opened a new chat for the next topic.
These tools were already very useful.
But the more I used them, the more another question interested me: what happens around the model to turn an answer into real assistance?
How does a conversation keep its context? What should be remembered? What can be forgotten? How does an assistant use a tool, consult a source or take an action without going beyond what it was allowed to do?
L'Alter was born from that curiosity.
Looking behind the interface
L'Alter was not born from the pretension of building a new model to rival those of OpenAI or Anthropic.
That is not the point.
It relies on existing models. The work lies in everything around them: the context they are given, the memory that is kept, the tools they are entrusted with and the limits they are held to.
A single answer can seem intelligent.
An assistant must do more. It must understand what is useful right now, find what it was told before and know when to use a tool rather than produce text.
It must also know what it is not allowed to do.
Learning before building
I did not start from a personal theory of agents.
I began by reading what already existed: context windows, context editing, conversation compaction, persistent memory, caching, agent-to-agent protocols and authorisation systems.
The Technical literature page gathers the work, documentation and standards that guide this construction.
It is not there to make the project look learned. It shows what our choices rest on and lets us go back to the sources when we need to question them.
We do not invent everything.
We try to understand, assemble and properly apply what already exists.
A very everyday need
Technical curiosity alone would not have made an application.
The need was much simpler: I wanted help organising what fills my days.
Tasks not to forget. Notes to keep. A piece of information to find. A reminder to set. A search to run without opening five apps.
I did not want to explain everything again in every new conversation.
So L'Alter started remembering what I told it and helping me act on it.
Gradually, it became my everyday assistant.
Tasks, notes and code
Today, I use L'Alter to organise my tasks, keep notes, search for information and find what I have already told it.
A code-oriented version also helps me work on the Fabrique's applications. It helps me review them, test them and track the problems that come up every day.
These are ordinary uses.
That is exactly what matters.
An assistant only becomes truly useful when it stops being an impressive demo and finds a place in the repeated tasks of every day.
An answer impresses once. Useful assistance has to hold up every day.
An assistant, not an oracle
L'Alter can be wrong.
It can misunderstand a request, miss a piece of information or produce an answer that needs checking.
It replaces neither the judgement nor the responsibility of the person using it.
That limit is part of the product. The aim is not to present an intelligence that knows everything, but a tool able to remember, search and act with you, within limits you can understand.
Trust does not come from the absence of error.
It comes from being able to understand what was used, check the result and keep the final say.
Then came Orbit
L'Alter changed how I use an assistant.
Orbit changes the very place where that assistance happens.
Today's assistants still live in a chat window, added to operating systems designed long before they existed.
The idea of using the Web as a real software platform is not new.
In 2007, before the App Store existed, Apple presented Web standards as a way to build applications for the iPhone. Steve Jobs said at the time:
"Our innovative approach, using Web 2.0-based standards, lets developers create amazing new applications…"
— Steve Jobs, 2007 (Apple's official announcement)
History then gave native applications an immense place. But the intuition remains: software can be distributed, used and updated straight from the Web.
Orbit takes up that idea on another scale.
What if the browser no longer just hosted applications, but became the operating system itself?
Orbit explores a new generation of operating system built right into the browser.
An environment where the assistant, the tools and the automations are no longer scattered across a multitude of applications, but brought together in one space.
The browser is no longer just for opening applications.
It becomes the system itself.
Orbit has only just been born. For now, just remember this: after building our assistant, we began rethinking the environment around it.
What comes next deserves its own episode.
The review
Get the next issues.
A few pieces a year, nothing else. No follow-ups, no promotions.