Autonomous AI agents

An agent does not hand you a draft. It does the thing: updates the record, sends the message, runs the workflow, and moves on to the next one.

That is why governance changes shape here. With an assistant, the risk is a wrong sentence. With an agent, the risk is a wrong action, taken without anyone reading first and possibly repeated many times before someone notices. Almost every common use of these tools acts without a person approving each step, and most can create, change or delete records in other systems.

We rate every use in this category at medium or high consequence and ask for a monthly look, the shortest rhythm on the site.

Governing an agent has less to do with checking outputs and more to do with bounding its reach. Give it the minimum access it needs, under its own identity rather than a person's login, so you can attribute what it did and switch it off. Keep a backup it cannot touch. Have a tested way to stop it fast. Check, independently, that what it reported doing actually happened, because an agent that fails quietly and carries on is the failure that hurts.

Two of these three tools can also talk to customers directly, which adds the customer-facing duties to the agent ones.

Open a tool page to see each agent's common uses and what it may do on its own.