What an agent actually is.

An agent owns a job. A chatbot answers a question.

That is the whole distinction, and everything else follows from it. Asking which invoices are overdue and getting "17" is an answer; the work is still yours. An agent checks every morning, decides which invoices merit a reminder, sends the ones it is allowed to send, escalates the rest and tells you what it did. The second thing needs far more machinery than the first, and almost all of that machinery is about limiting it.

The six things an agent needs

Something missing any of these is a script with a language model attached. The absences are usually identity and ownership, and those are the two that matter most when something goes wrong.

• An identity of its own, so you can tell what acted
• A purpose, written down, in the words of whoever asked for it
• Capabilities it was granted — not credentials it holds
• Limits on what it may decide without a person
• A named human who is accountable for it
• A record of what it did, and why

Why capabilities rather than credentials

The shortcut is to give an agent an API key with broad access and instruct it carefully. This fails in a specific way: the instruction is a request rather than a limit, nobody can tell afterwards which agent acted, and revoking the key breaks everything else using it. A capability — "read an invoice" — is narrow enough to review, attributable to one agent, and revocable on its own.

Autonomy is a range, not a switch

The useful question is never whether an agent should be autonomous. It is which specific decisions it may make alone. Sending a payment reminder at day nine: yes. Agreeing a three-month payment plan: no. Most agents that fail in practice were configured as though that were one decision.

The first agent should be boring

A morning briefing that reads your systems and emails four managers is a good first agent. It takes no consequential action, it costs almost nothing, and it teaches you what the run timeline looks like when something is wrong. Teams that start with the highest-value process spend their first fortnight nervous rather than learning.

What an agent is bad at

Anything needing judgement you would be uncomfortable explaining afterwards. Anything where being confidently wrong is expensive. Anything that must happen within a second — this is work measured in minutes. And anything genuinely fixed in sequence, which an ordinary automation tool does more cheaply and more predictably.