Agent Frameworks

LangGraph, PydanticAI, CrewAI — they hand you the loop, state, retries, and tracing as reusable parts. They add structure, not new powers.

Building with AI · Lesson 21 · 11 min read

You’ve now hand-built an agent loop (L19) and bolted on guardrails (L20). Starting your next, bigger project, you notice you’re about to re-write the same plumbing — the loop, tool wiring, retries, logging, state — all over again. And a teammate wants two agents to collaborate. Re-coding this every time is wasteful and error-prone. Agent frameworks exist so you don’t have to. But which one, and what do they actually give you that you couldn’t already build?

A framework is reusable plumbing, not new powers

Everything from the last two lessons — the loop, tool wiring, state between steps, retries, error handling, tracing, coordinating multiple agents — is plumbing you’d otherwise rebuild each time. A framework packages it as tested building blocks with sensible conventions and built-in observability. The point to hold onto: a framework gives you no capability the model itself lacks. Same model, same tools, same loop you already understand — just organized for you, so you write less boilerplate and get logging and structure for free.

The same weather agent, two ways:
• Hand-rolled: you write the loop, the retry logic, the step budget, and the logging yourself.
• Framework: you declare the tools and the goal; the framework runs the loop, handles retries, enforces a step limit, and traces every step for you.
Identical behavior — the framework just absorbed the boilerplate you’d otherwise own.

Different frameworks, different emphases

Popular frameworks each optimize for a different need:
• LangGraph models an agent as a graph / state machine — explicit nodes and edges for control flow. Great when the flow is complex or cyclic, and for checkpointing and human-in-the-loop pauses.
• PydanticAI emphasizes type safety: structured outputs validated against a schema, plus a clean, Pythonic, dependency-injected style. Great when you need validated, typed results.
• CrewAI focuses on role-based multi-agent “crews”: define agents with roles and tasks and have them collaborate. Great for quickly assembling a team of cooperating agents.
The pattern: each trades some raw flexibility for structure around one particular need.

When to use one — and when not

Don’t reach for a framework reflexively. For a simple agent — a few steps, one or two tools — a plain loop plus your own guards is fine and easier to debug. Frameworks earn their keep as complexity grows: many tools, branching control flow, multiple agents, production-grade tracing. Choose by the need (explicit control flow → a graph framework; validated outputs → a typed one; agent teams → a crew one), not by hype, and remember you can always drop back to plain code. Honest caveat: frameworks add a learning curve and abstraction that can hide what’s happening — which is exactly why understanding the underlying loop (you now do) is what lets you wield one well, or skip it on purpose.

Matching four needs to the right choice:
1. A support agent that must branch and PAUSE for human approval before issuing a refund → a graph/state-machine framework (LangGraph) for explicit flow + human-in-the-loop.
2. An agent that must always return a strictly-typed, schema-validated record → a type-safe framework (PydanticAI).
3. A “research team”: a planner that delegates to a writer and a fact-checker → a role-based crew (CrewAI).
4. A 3-step script that calls one tool → no framework; a plain loop with your own guards.
Same models and tools throughout — the choice is purely about structure fitting the job.