AI
Building AI agents you dare to run in production
An agent doing something clever in a demo has become ordinary. An agentic workflow your business process can lean on is a different trade. Three properties make the difference.
Restartable, scalable, auditable
Restartable. Everything long-running goes down eventually: a server restarts, an API disappears for a while, a model returns something unusable. A production-grade workflow knows where it was and continues from there, without duplicate work and without loss. Not start over and hope, but resume and prove.
Scalable. One agent doing one task is a demo. A hundred tasks at once, with queues, priorities and the rate limits of external systems, is reality. The workflow has to be built for that, because bolting scalability on afterwards means rebuilding.
Auditable. When an agent acted instead of a person, the question "what exactly happened, and why" counts double. Every step, every decision and every human approval belongs on record, in a form you can show a customer or an auditor. Without that record, an agent has no place in a business process.
Where the human belongs
The question is not whether a human belongs in the loop, but where. Our answer: at the points where a mistake cannot be undone or costs money and reputation. An agent may prepare, categorise and stage an e-mail; the commitment to the customer is signed by a person. A good workflow makes that approval moment explicit, instead of trusting that someone happens to be watching.
Waiting is a core skill
Business processes wait constantly: for a supplier's answer, for a batch run, for a system that will be back tomorrow. An agent that only works while everything answers instantly is useless for those processes. Our workflows can wait days for a slow or asynchronous system, lose nothing in the meantime, and pick the thread back up the moment the answer arrives. The same property that makes our integrations reliable makes agents production-grade.
Running in production, at our place
In TenderScan.nl, workflows process thousands of tenders every week into relevant reports for users. On Supertaak, our workflow platform where execution and accountability are equally central, an AI-driven e-mail automation runs at a client, handling daily work with a person at the decision points. Both run in production; both are built to exactly the three properties above.
How we build with AI agents ourselves, including what goes wrong, is in the article AI agents in real projects.
An agentic workflow for your process?
Tell us which process you have in mind, and we will say honestly whether an agent belongs there, and where the human should stay.