Agentic Patterns
Most useful agents fall into a handful of shapes. A pattern is a shape: what wakes the agent up, what it does, and where a person stays in charge. Knowing the shape of what you want to build makes the first design decision easy, and it tells you what can go wrong.
This is reference. The Idea Gallery is where ideas live, and the guided path teaches the parts you need inside each step.
The five shapes
| Pattern | Starts when | What it does |
|---|---|---|
| Watcher | A schedule or an event | A schedule or an event wakes it up; it notices something and tells a human. |
| Operator | A request from a person | It acts on real things, and a human approves before anything irreversible happens. |
| Investigator | A case to explain | It correlates signals across sources and hands a human a case file, not a guess. |
| Concierge | A person asking | A customer or co-worker talks to it; it answers from grounded sources and knows when to hand off. |
| Coordinator | A question that spans several systems | It orchestrates other agents or tools over MCP and makes a go/no-go call with a human. |
How to choose
- Does a person ask, or does something just happen? A person asking is a Concierge or an Operator. Something happening on its own is a Watcher.
- Does it only tell, or does it change something? Only telling is a Watcher, an Investigator or a Coordinator. Changing something is an Operator, and it needs a human approval.
- Is the evidence in one place or many? Many places point to an Investigator (one case) or a Coordinator (one decision across systems).
Real agents often mix two: a Watcher that hands what it finds to an Investigator, or a Concierge that hands a request to an Operator. Start with the one that matches the first thing your agent must do.