// ai

AI agent

An AI agent is a system that does not just produce text: it calls tools — APIs, databases, calendars — keeps state across steps, and pursues an outcome, like booking a slot, processing a refund or triaging tickets. A chatbot advises; an agent acts, which is why its failures are measured in consequences, not word quality.

Agents matter where the value is in the transaction, not the conversation: systems that check real availability before booking, that process the refund end to end, that read an inbox and route each message. The tell is a checklist — if a human performs a fixed sequence of reads and writes across systems, an agent can probably do it. The architecture differs from a chatbot by three additions: tools, state, and a control loop of plan, act, observe, retry.

The cost and risk shift with the capability. A chatbot conversation costs fractions of a cent; an agent task — plan, five tool calls, retries — costs ten to fifty times more, and the bill grows with the length of the task. More important is the permissions problem: the moment software can write data, you need scoped keys (never admin credentials), idempotent actions, a dry-run mode and a log of every step. Honest deployments we have seen land between a few cents and a few euros per completed task.