AI Agents Replaced an IT Department (Part 1 of 2)

AI Agents Replaced an IT Department: How It Was Built

Amadeni AI's Delia Hallberg on multi-agent coding and hallucination design

Two founders, the output of four or more people, and a set of AI agents that Delia Hallberg describes as a coherent system standing in for an entire IT department. She is also emphatic that it is not vibe coding — and in this conversation she says precisely what the difference is.

Hallberg co-founded Amadeni AI, one of the newer German startups, building AI-supported custom software for Mittelstand Germany. Part one of two is about the engineering and the diagnosis underneath it: what the agents actually own, where the human stays, and how you design a system so a language model is never forced to invent an answer.

Top signals from this episode:

1. The agent split at Amadeni. Roughly 80 percent of marketing and content runs through agents holding the strategy, plan and pipeline. Sector research happens before a first client call, because the firm works across industries and regularly starts in one it does not know. A sales workflow hands a converted lead to delivery with a project outline already written. And multiple coding agents carry the largest gain. The pre-AI equivalent was one generalist plus at least three developers.

2. The boundary test, and the single most useful sentence here for anyone running agents in production: you understand your agents well enough if you could do the work yourself should they stop tomorrow. Agents replace capacity, not competence.

3. Anti-hallucination architecture at the Ostdeutscher Bankenverband, whose member banks make decisions on the information it publishes. Not a zero-hallucination claim — a risk-lowering design: a human in the loop at every step, different unconnected models each given one small concrete task, and traceability from every output item back to its original source. Her reasoning is that a model invents when it lacks information but still has to answer, so shrinking the task shrinks the pressure to invent.

4. The diagnostic that decides whether automation is even possible: if nobody can describe a process without naming a colleague, it is a set of personal habits, and automating it encodes a staffing arrangement rather than a workflow.

Also covered: why the specific danger of agent systems is that plausible output lets you fool yourself for a long time before noticing; documenting agent architectures the way you document software architectures, so they scale instead o