Agent frameworks: do you need one?

The optionsCopy Anchor Link

LangGraph models an agent as a graph of steps with persistent state. Pydantic AI builds typed agents around validated outputs. The OpenAI Agents SDK and Claude Agent SDK provide the loop, tools, handoffs and guardrails from each lab. smolagents from Hugging Face is a small library where agents write and run code. CrewAI and AutoGen focus on multiple cooperating agents. LlamaIndex and Haystack lean toward retrieval and document pipelines. The OpenAI Agents SDK (with AgentKit for visual building), Google's Agent Development Kit (ADK) and Vertex AI Agent Builder are the labs' own offerings.

Do you need one?Copy Anchor Link

You can write the core loop in under 100 lines. A framework is worth it when you need durable state, human approval steps, tracing or many integrations. It is a cost when it hides prompts or makes behavior hard to inspect. A sound path is to write the loop yourself once to learn it, then choose a framework with that understanding.

A worked example: when a framework earns its placeCopy Anchor Link

You need an agent that can pause for human approval, survive a server restart and resume from the same step. Building that yourself takes weeks. A framework with checkpoints gives it to you. For a script that calls a model twice, a framework is more weight than help.

How to evaluate oneCopy Anchor Link

Can you see the exact prompts it sends? Can you swap models? How does it handle errors and retries? Who maintains it and how often does it break? Try it on a small real task before committing.

What a framework does for youCopy Anchor Link

Every agent has the same loop: call the model, check for a tool call, run it, add the result and repeat until done. A framework wraps that loop and adds three things: saved state, retries and tracing.

Saved state is the part beginners miss. Imagine a six-step task that pauses for approval at step 4. The framework writes a checkpoint after each step. If the server restarts, it resumes at step 4 instead of step 1. Doing this by hand means handling storage, repeated steps and half-finished tool calls.

A second example: the framework gets in the wayCopy Anchor Link

You want to change one sentence of the system prompt. Some frameworks add their own instructions around yours and builds the final prompt several layers down. Your edit does nothing visible. To confirm, log the raw request sent to the provider, or use the framework's tracing, and read what the model actually received. If you cannot do either, treat that as a warning sign.

Trade-offsCopy Anchor Link

Benefits: less code for state, approvals and tracing, and a shared structure for a team. Costs: a new set of ideas to learn, upgrades that can break your code, and abstractions that leak when something goes wrong. Moving off a framework later is work, because its concepts spread through your code.

Real mechanism versus marketing: nothing in a framework makes the model smarter. It organizes the same API calls.

Try it yourselfCopy Anchor Link

1. Write the loop yourself in about 30 lines, with one tool and a step cap.
2. Rebuild the same agent in a framework you are curious about.
3. Compare line counts.
4. Check whether you can print the exact prompt each version sends.
5. Break a tool on purpose and see which version gives the clearer error.

Expect the framework version to be shorter and the hand-written one to be easier to inspect.

Common mistakeCopy Anchor Link

Choosing by popularity alone.