Skip to content

Agent patterns

Four shapes that work.

Across the jobs businesses actually hand to agents, almost everything successful is one of four shapes. Recognising which one you are building tells you where the approval belongs before you have written anything.

Chase — something is overdue and nobody followed up

Overdue invoices, unanswered leads, stalled onboarding steps, documents a project is waiting on. The agent finds the items past a threshold, acts on the ones inside a safe window, and escalates the ones outside it. The approval belongs at the threshold where the contact stops being routine and starts being delicate — which is usually the moment somebody replies.

Triage — high volume, mostly answerable, occasionally not

Support questions, enquiries, inbound requests. The agent answers what it can verify from approved knowledge and hands over everything else with the context already gathered. The failure mode is an agent that answers when it should not, so the approval belongs on any commitment about the future and any goodwill gesture.

Brief — assemble what a person needs to decide

Morning briefings, renewal packs, exception summaries, report preparation. The agent reads several systems and writes one thing. These take no consequential action at all, which makes them the cheapest and safest place to start, and the best way to learn what your data actually looks like.

Watch — notice the thing that changed quietly

Campaign performance, account health, page performance, obligations approaching a date. The agent looks at the same things daily and says something only when a threshold is crossed. The discipline here is silence: a watching agent that reports every day trains people to ignore it.

The two that keep failing

An agent asked to own an outcome rather than a job — "increase revenue" — has no threshold, no boundary and no way to be wrong, which means it cannot be reviewed either. And an agent given a long fixed sequence, which is an automation dressed up: it costs more, behaves less predictably, and an automation tool would have done it properly.

  • Give it a job with a threshold, not an outcome with a hope
  • If the steps never vary, use an automation tool instead
  • If you cannot say what a wrong action would look like, it is not ready

Give AI a job.

Build an agent, choose what it can access, define what it may do, and deploy it to work safely across your business.

No code required · Full SDKs and API · Stop any agent instantly · [email protected]