Agentic Design Patterns: The Shapes Every Coding Agent Reuses

Short summary

Agentic design patterns are named control structures for organizing LLM calls and tools into five reusable shapes: prompt chaining, routing, parallelization, and orchestrator-workers. The core decision rule: pick the simplest pattern solving your task; agentic systems cost 4-15× more tokens but unlock performance on open-ended work. Verification loops (gather-act-verify-repeat) replace model guessing with ground-truth feedback.

• •Five reusable patterns (prompt chaining, routing, parallelization, orchestrator-workers) map to specific task types
• •Escalation rule: start simple, only add complexity when a specific failure mode forces it (4-15× token cost)
• •Verification loops make agents distinct from scripts—gather context, take action, verify result, repeat until green

Generated with AI, which can make mistakes.