Skip to content

Pick a framework

Not sure which one to choose? Use this table.

If you want… Pick Output Best for
Multiple cooperating roles, prod-grade scaffolding CrewAI Python + YAML Most teams. Familiar Python project layout.
Explicit state and graph control LangGraph Python Pipelines, long-running flows, conditional routing.
To deploy inside IBM watsonx Orchestrate WatsonX Orchestrate YAML Enterprise IBM stacks.
Event-driven mini-crews CrewAI Flow Python Multi-step pipelines with branching.
A single-file reasoning loop ReAct Python Tool-calling agents, demos, embedded use.

CrewAI

Generates a complete project:

project/
  config/
    agents.yaml     # role, goal, backstory
    tasks.yaml      # description, expected_output, agent
  src/
    crew.py         # @CrewBase with @agent, @task, @crew
    main.py
    tools/          # tools picked from the catalogue
  tests/test_smoke.py
  pyproject.toml    # crewai>=1.12
  README.md  .env.example  .gitignore

agent-generator "Research team with a researcher and a writer" \
  -f crewai -o team/

Modes: code_only, yaml_only, code_and_yaml.

LangGraph

A single graph.py with:

  • TypedDict state schema
  • one node function per task
  • edges from task dependencies
  • a START entry point
  • graph.compile() + app.invoke()
agent-generator "Extract, transform, load pipeline" \
  -f langgraph -o pipeline.py

Mode: code_only.

WatsonX Orchestrate

Emits the ADK YAML:

spec_version: v1
kind: native
name: agent-name
description: What the agent does
instructions: |
  Detailed instructions from task goals
llm: watsonx/meta-llama/llama-3-3-70b-instruct
tools: [tool_name]
knowledge_base: []

agent-generator "Customer support assistant" \
  -f watsonx_orchestrate -o support.yaml
orchestrate agents import -f support.yaml

Mode: yaml_only.

CrewAI Flow

Single Python file with:

  • FlowState (Pydantic) for shared state
  • WorkflowFlow(Flow[FlowState])
  • @start() and @listen() decorators
  • Each step spawns a mini-Crew
agent-generator "Content pipeline: research, write, edit" \
  -f crewai_flow -o content.py

Mode: code_only.

ReAct

Single file with:

  • Tool registry via @register_tool
  • think() / act() functions
  • react_loop() with a MAX_ITERATIONS guard
  • Per-task runners; built-in search and calculate
agent-generator "Code review bot" -f react -o reviewer.py

Mode: code_only.

Capability matrix

Feature CrewAI LangGraph WatsonX Flow ReAct
Code output ✓ ✓ ✓ ✓
YAML output ✓ ✓
Code + YAML ✓
Tool templates ✓ ✓ ✓ ✓
MCP wrapper ✓ ✓ ✓ ✓

Next: Architecture · Platform overview