Skip to main content

How Livalos works

We sell agentic AI and we run on it. Here you can read how that works in practice, without jargon and with real figures.

SecondBrain, the memory

Everything Livalos knows is in 1 system: SecondBrain. It is a collection of notes and tasks in text files. Customers, agreements, decisions, quotes and tasks. Every agent reads the same source, so nobody works from an outdated picture.

Every piece of work has a task. If there is no task, the work is invisible, and invisible work gets lost. An agent creates the task, works on it, notes what it did and closes it. That way Maarten can see at any moment what is going on and why.

The agent team

Each agent has 1 role. One agent manages the website and support. A second maintains the platform. A third builds the automation. A fourth follows up on sales. A conductor divides the work and pings an agent when something is waiting.

We started in June with 26 roles and at the peak there were 32. That turned out to be too many, because more agents meant more tasks, more notifications and more costs. Three months later 9 were working. The lesson: start with few roles and let each do 1 thing well.

The Telegram tap

An agent can do a lot, but not everything. Anything that goes out gets approval from Maarten. Every email to a customer first arrives as a message in Telegram, with the text included. Maarten taps approve or rejects it. Only after the tap does the email go out.

That is the rule: agents execute, Maarten decides. Whatever costs money, goes to customers or goes live gets a human.

The IssueFixer

A bug on our platform becomes an issue in GitLab, in English. Our IssueFixer picks up any issue with the label `issuefixerbot`. The bot pulls in the code, has Claude write the fix and creates a merge request. A reviewer checks that request and tests it on a real test site before it goes live.

The result in 2026: 741 merged changes on the platform, 160 in September and 141 in October. On the dashboard that was 259. Read the full case.

The release rhythm

Approved changes pile up and go to customer sites in a release. In 2026 we made 25 releases. A release goes first to a test environment, then to a small group of sites, and only then to all sites. That way we catch errors before they reach your site.

The MCP, the door for your agents

Our own agents work with the same tools we give to customers. The Livalos MCP gives an AI assistant safe access to a website, the CRM and quotes, with 213 tools behind a login (OAuth) and with limits per hour. Read the case about the MCP.

What this means for you

You see how an organisation works where agents do the routine work and a human makes the decisions. And you can do it yourself. We teach it in our training courses, we map it out with you in an advisory project, or you use the platform we run on ourselves.

One email a month

What is new at Livalos: free tools, apps and trainings. You confirm with a link in your email, unsubscribe in one click.