Kinds of Frameworks

Three styles

Graph-based tools such as LangGraph model your agent as nodes and edges with explicit state, good for controllable workflows. Agent SDKs from model providers (for example the OpenAI Agents SDK or the Claude Agent SDK) give a ready-made loop, tools and handoffs with little code. Role-based multi-agent libraries such as CrewAI let you describe a team of agents with roles and tasks. Features and names change quickly, so check current documentation before choosing.

The same idea as a graph

This pseudocode shows why graphs help: every step and every branch is explicit, so you can pause, inspect or replay it.

START -> plan -> call_tools -> check
                        ^           |
                        |   not done|
                        +-----------+
                              done  -> END

state = { messages, plan, results, step_count }

Features change fast

Framework names, APIs and recommended patterns shift every few months. Pin the version you use, read the changelog before upgrading, and keep your own tests so an upgrade cannot silently change behaviour.

Quick check: What is the main appeal of a graph-based agent framework?

• Explicit, controllable steps and state
• It needs no model
• It is always faster
• It avoids all bugs