Agents that do a job, not a demo. An AI agent gets a goal, uses tools and takes several steps on its own: it reads an email, looks up the customer, drafts a reply and files the case. I build agents that do one well-defined job in your business, and that you can see, measure and steer.
Book an intro callA chatbot answers questions. An agent does work. It fetches data from your systems, makes simple decisions by your rules and hands over to a person when something is unclear.
That makes agents useful for tasks where someone copies information between systems today, but also riskier. So most of the work is about limits: what the agent may read, what it may change and what needs approval.
I use a crew of AI agents every day myself, in a harness I built where every request, tool call and cost is visible while it happens. You get the same visibility in what I build for you.
Every agent logs what it read, which tools it used and what it decided. You see the cost per case and how often the agent gets it right, measured against real examples from your business.
The agent gets only the permissions the task needs. Steps that move money, change contracts or reach customers can go through human approval. Data that must stay in the EU is processed there, or by a model running on your own server.
I pick the model for the task: a large model from Anthropic or OpenAI when judgement matters, a smaller and cheaper one when the task is simple, and open models on your own server when data cannot leave the building. The agent is built so the model can be swapped without redoing the rest.
The agents are written in TypeScript and run in your environment or on a server I operate, with the same monitoring as the rest of your systems.
A typical first engagement is an agent for a single kind of case, such as questions about deliveries or invoices. It can usually run in shadow mode within a couple of weeks, and go live once the numbers hold.
Read more about how I work as an AI consultant in Gothenburg.
My own agent harness, where a crew of AI agents works in my projects. Every request, tool call, file change, cost and approval is visible while it happens. It is how I build, and you get the same visibility in what I build for you.
Read the casePrivate project, in daily use
One scored audit of everything you ship: code, sites, apps, APIs, containers, infrastructure, cloud and mobile builds. It runs from the terminal, in CI or from an AI agent over MCP.
A chatbot answers questions in a conversation. An agent has a goal and tools: it can look up data, update a system and take several steps on its own, within the limits you set.
Yes, when the agent has only the permissions the task needs, everything is logged and risky steps need approval. That is how I build, and everything can be reviewed afterwards.
It depends on the model and the volume. I measure the cost per case already in shadow mode, so you know what the agent costs to run before it takes over, and I use a smaller model where it is enough.
Yes. Agents work through APIs and MCP with CRMs, business systems, inboxes and document archives. Where there is no API, I build an integration first.