Plan-and-Execute & Routing

Two patterns for complex agents: upfront planning that decomposes tasks into steps, and routing that dispatches to specialized sub-agents.

Quick Reference

→Plan-and-Execute: a planner LLM creates a step-by-step plan, then an executor agent runs each step sequentially
→Store the plan in graph state as a list of steps with status tracking (pending, running, done, failed)
→Routing: a classifier node reads the input and routes to one of N specialized sub-graphs based on intent
→Use structured output for the router (intent: enum, confidence: float) to make routing deterministic
→Combine both patterns: route to the right domain, then plan-and-execute within that domain

Plan-and-Execute Pattern

When to use

Plan-and-Execute shines when the task has 4+ steps that depend on each other, and the agent needs to track progress across those steps. Examples: research reports, data analysis, multi-step API workflows.

Plan-and-execute: planner decomposes the task, executor runs steps, replanner adjusts until done

Plan-and-Execute agent with dynamic re-planning

Routing Pattern

Routing pattern: an LLM router classifies intent and dispatches to the best specialized handler

Combining Plan-and-Execute with Routing

The most powerful agents combine both patterns: route to the right domain first, then plan-and-execute within that domain. This gives you specialized handling without monolithic complexity:

Sign in to read this article

This is a premium article. Sign in with your Google account to continue.