Artificial intelligence system that does not merely produce a textual output but performs acts in the world, calling tools, writing into systems and triggering operations without an operator reviewing each decision.
An autonomous agent differs from a language model in four technical properties, each carrying an insurance consequence. Tool access gives it effect in the outside world, which converts an error of content into physical or financial damage. The execution loop, by which the agent observes the result of its action and proceeds, allows an initial error to propagate across a long sequence before any detection. Persistence, meaning the capacity to operate over time without prompting, removes the synchronisation between activity and supervision. Composition, where several agents call one another, creates dependencies neither the deployer nor the insurer maps in full. For underwriting the decisive consequence is that the unit of exposure changes nature: it is no longer measured by the insured's turnover or headcount but by the number of actions executed and the privileges granted. A singularity follows that the trade had never met, claims frequency stops being a property of the world the insured undergoes and becomes a property of configuration, alterable within minutes without endorsement or notice to the insurer.
A fifty-person firm running a fleet of commercial agents may commit more legal acts per day than a group of five thousand people. A premium calculated on headcount then misses the risk entirely.
agent IA, agent agentique, système agentique, agentic system, AI agent