Agents that do a job, inside limits you set

A chatbot answers; an agent acts. A chatbot tells your customer the returns policy. An agent looks up their order, checks eligibility, creates the return and books the collection. The second one needs permissions, limits and an audit trail, which is most of the work.

Four things, and none of them is trust in the model: tools that can only perform narrow validated operations, approval gates on anything irreversible, hard limits on steps and spend enforced in code, and an audit log so a mistake is diagnosable. We design assuming the model will occasionally be wrong, because it will.

For narrow, well-defined tasks with a human on irreversible steps, yes — we build and run these. For open-ended autonomy over important systems, not yet, and we will say so rather than sell you a pilot that quietly gets shelved.

You constrain it rather than eliminate it. Facts come from your systems through tools instead of from the model’s memory, outputs are validated against a schema before use, and anything unverifiable becomes an escalation rather than an action. Any supplier promising zero hallucination does not understand the technology.

Whichever benchmarks best on your evaluation set, behind an adapter so it can be changed. In practice we most often use Mistral or Claude for tool use. The important thing is that the choice is measured on your tasks and reversible, because the leaderboard changes every few months.