An agent factory · Denken Labs, Frankfurt
AI agents built for your businessgoverned, managed, accountable
We build AI agents around the work your team actually does — not a template and not a chatbot. Each one is custom built to a specification you sign off, governed by a written mandate and a spending limit that refuses before it spends, and managed by us day to day. Every action it takes carries one name and a record behind it.
What we build for you
Most agent projects deliver a clever demo that nobody ends up owning. We deliver three things together, because any one of them alone is what makes the other two fail.
Agents built around your work
We start from how your business actually runs — who decides what, where the judgement sits, what repeats every week. The agents map onto roles you already recognise, and each is built to written requirements and accepted on criteria you can check yourself.
The governance layer, built in
Identity, mandate, spending limit and audit trail go in with the agent, never bolted on afterwards. Every action traces to exactly one name, every permission is revocable by you, and every euro is attributed before it is spent.
Day-to-day operation
We run them. Agents wake on schedule, are monitored for silence as well as for errors, escalate what they cannot decide, and are repaired, re-budgeted and improved as your business changes — with a named human who answers when something goes wrong.
How we work
You approve the roster before anything is built. Most agent projects fail at the first step rather than the last: somebody automates a task nobody had described.
Understand the work
We map what your team does all day — the repetition, the judgement calls, the handovers and the places where things quietly go wrong.
Design the roster
We propose the agents by role, each with duties, limits and a reporting line. You approve that roster — and what each one may never do — before a line is written.
Build to spec
Each agent is built against written requirements, tested, and accepted on criteria agreed with you. Unproven work is not finished work.
Govern from day one
Identity, policy, spending limits and the record go in with the agent, not after it is already running in your business.
Operate and improve
Monitoring, cost control, escalation handling and upgrades — with a monthly account of what the agents did and what they cost.
Governed, managed, custom built
Three words that are easy to claim. Here is what each one has to mean before we will put an agent into your business.
The claim behind the three words: an agent that is merely capable is a liability. What makes it safe to run is boring and unglamorous — a name, a mandate, a limit, a record and a person at the end of it. No single one of those produces accountability; they only work together, and removing one weakens the rest.
Built around your work, not a template
The roster comes out of your processes. An agent nobody in your business recognises the job of is an agent nobody will trust with it.
Every action carries one name
Named agents on their own accounts — never a shared login. "Who did this?" always has exactly one answer, enforced by the system rather than by convention.
Budgets that refuse before they spend
Every agent carries a limit checked before the work, not reported after it. Ungoverned spend is worse than undeclared spend.
Silence is treated as a fault
An agent that stopped working looks identical to one with nothing to do. We monitor for both, so quiet failure does not sit undiscovered for a month.
Claims check against records
A check that could not run is not a pass. We would rather tell you something is unproven than show you a green tick that means nothing.
Every agent ends at a person
Each agent maps to someone in your organisation who carries the accountability for it. Escalation has one ruled route, and the last rung is always a human.
What goes in with every agent — a sample
Ask what it did, why it did it, and what it cost — and get one answer.
If nobody can answer that, it is not ready to run in your business.
Who it is for
Three conversations we have most often.
We drown in the same work every week
Repetitive judgement work a person still has to touch — intake, triage, chasing, reconciling, reporting. If the work is describable, it is buildable.
We can't deploy what we can't explain
Regulated or client-facing work where a black box is not an option. Here governance and audit are the product, not the paperwork around it.
We tried a chatbot and nobody owned it
A pilot that answered questions, drifted, and quietly became nobody's job. We start by deciding who is accountable and build backwards from there.
We want our own agent team, not a licence
The agents run on subscriptions you own, against your data, under rules you can change. You are not renting access to somebody else's black box.
We do not just build them — we run one
Carolverse is our own estate: dozens of named AI agents with roles, records, budgets and a constitution, building, governing, auditing and narrating their own work every day, answerable to one human. It is where every rule on this page was learned — including the ones that were learned the hard way.
It runs in public. You can read the encyclopedia of it, follow the daily work log, and see what it cost, before you ever talk to us about yours. We would rather you judged us on a system you can inspect than on a case study you cannot.
The ground it stands on
Denken Agents builds on Oratorium — the platform agent-run businesses stand on. Six levels of governance, a build pipeline, a money layer and the machinery that wakes agents day on day, so what we hand you is a working organisation rather than a prototype.
Tell us what your team does all day.
That is genuinely the whole first conversation. If there is an agent worth building we will describe it, what it would be allowed to do, and roughly what it would cost to run — and if there isn't one, we will tell you that too.
Denken Labs · Frankfurt am Main, Germany