Foundations · start here

Most agent systems are one of a few shapes.

Learn the shapes and you can read, design, and explain almost any workflow. This atlas maps 23 patterns — from a single agent thinking in steps, to teams that delegate, debate, and hand off. Each one carries a topology glyph, a prompt you forward to your own agent, and the part that matters most: where you, the human, stay in the loop.

Browse single-agent Jump to multi-agent

The one rule

Reach for the simplest pattern that works. A workflow on fixed rails is predictable and cheap; an autonomous agent is flexible but harder to control. More agents is not better — add coordination only when a measurable problem demands it.

Workflow or agent?

Workflow

You design the path. The model fills in the steps.

Predefined sequence of LLM steps and tool calls.
Predictable, testable, easy to debug.
Best when the route is known up front.

Chaining · routing · sectioning · evaluator–optimizer

Agent

The model directs its own process and tools.

Decides what to do next, loops until done.
Flexible on open-ended tasks; less predictable.
Best when the path can't be known in advance.

ReAct · orchestrator–workers · swarm · monitoring

How to read a pattern

The glyph is the topology.Circles are agents, the diamond routes, rounded shapes are shared stores or boards, squares are tools. Indigo edges are the flow of work.

N agentsCore

Agents & level.How many agents the pattern needs, and how much machinery it takes to run well — Starter, Core, or Advanced.

Where you stay in the loop.Every pattern names the human's seat — what you own, where your judgment matters, and what should never happen without your say-so. The shape connects the agents; this is where you connect to it.

The forwardable prompt.Each pattern ends in a copy-ready brief. Fill the placeholders, paste it to your agent, and it sets the workflow up with you.

Built as a living reference — fork it, extend the taxonomy, and share it with anyone setting up their own agents.