Four Controls That Turn an AI Agent From a Demo Into a System

Marius Andronie, who runs the agency Devaland, starts from the difference between a chatbot and an agent: one answers, the other acts. It searches, fills in forms and sends replies. So an agent needs rules before it needs features, and he boils them down to four controls, each framed as a question to ask a vendor.

Access should be scoped to the task, time-limited where possible and fully logged, rather than broad and standing because that is convenient. Output should get a second check before it leaves the company: a person who approves it, or a separate evaluator agent that scores the work, as in one London firm’s executor-and-evaluator pairing. The agent needs rules, set in advance, for when it does not know: say there is no source, ask a person when unsure, escalate when money or an upset customer is involved, and test it with oddly phrased requests, since tidy ones nearly always work. And every run should be logged (input, sources, output) so any past answer can be traced back.

He draws the four from two 2026 Sifted guides on agentic AI and closes on his own firm’s practice: answers cite a source for each claim, and say so when there is none.