29 · Agent Design Patterns

Different problems need different patterns. This catalogs agent design patterns and when to use each, so you start simple and scale as needed.

How it works

1. 1Sequential (chain): step-by-step tasks where each step depends on the previous.
2. 2Parallel: independent subtasks run at once, then merge.
3. 3Routing (conditional): different requests take different paths.
4. 4Supervisor (orchestrator): a central planner delegates to specialists.
5. 5Collaborative, reflection/evaluator, autonomous and hierarchical patterns for harder problems.

Key points

• Use a selection guide: is the flow linear? are subtasks independent? do requests differ?
• Patterns can be combined (e.g. routing + supervisor + reflection).
• Implement in LangGraph on the AWS stack.
• Start with the simplest pattern that works, then iterate.

Why it matters

Matching the pattern to the problem avoids both over- and under-engineering. The right pattern makes an agent reliable, efficient and easier to maintain.

Frequently asked questions

How do I pick a pattern?

Ask whether the flow is linear, whether subtasks are independent, and whether requests need different paths — then match the pattern.

Can patterns be combined?

Yes — real systems often mix routing, supervisor and reflection patterns.