Multi-Agent Orchestration with CrewAI and LangGraph

The Multi-Agent Problem

Building production multi-agent systems means choosing an orchestration framework early, and that choice has deep implications. CrewAI excels at role-based sequential workflows. LangGraph provides fine-grained control over stateful computation graphs. AutoGen offers conversational patterns. Each has trade-offs.

Agentic Assistants solves this with a uniform adapter interface -- you define your agents and tasks once, and the framework handles the translation to whichever backend you choose.

The Adapter Interface

Every agent framework adapter implements a common protocol:

class AgentAdapter(Protocol):
    def create_agent(self, config: AgentConfig) -> Any: ...
    def create_task(self, config: TaskConfig) -> Any: ...
    def execute(self, agents: list, tasks: list, **kwargs) -> ExecutionResult: ...

This means switching from CrewAI to LangGraph is a one-line config change rather than a rewrite of your agent definitions.

CrewAI Adapter

The CrewAI adapter (adapters/crewai_adapter.py) maps agent configs to CrewAI's Agent and Task objects, then assembles them into a Crew. It supports CrewAI's sequential and hierarchical process types, tool binding, and memory.

A typical workflow definition looks like:

framework: crewai
process: sequential
agents:
  - name: researcher
    role: "Senior Research Analyst"
    goal: "Find comprehensive data on the topic"
    tools: [web_search, document_reader]
  - name: writer
    role: "Technical Writer"
    goal: "Produce a clear, detailed report"
    tools: [text_formatter]

The adapter translates this YAML into CrewAI objects, injects the appropriate LLM (via Ollama), and executes the crew.

LangGraph Adapter

The LangGraph adapter (adapters/langgraph_adapter.py) takes a different approach. Instead of role-based crews, it constructs a StateGraph where each node is an agent or tool invocation, and edges define the control flow.

This is more powerful for complex workflows -- conditional branching, parallel execution, human-in-the-loop checkpoints, and cyclic graphs are all natively supported.

graph = StateGraph(AgentState)
graph.add_node("research", research_agent)
graph.add_node("analyze", analysis_agent)
graph.add_node("decide", routing_function)

graph.add_edge("research", "decide")
graph.add_conditional_edges("decide", route_logic, {
    "needs_more": "research",
    "ready": "analyze"
})

When to Use Which

CrewAI works best for straightforward, role-based pipelines: research -> analyze -> write. The mental model is a team of specialists with a clear chain of responsibility.

LangGraph shines when you need complex control flow: retry loops, conditional routing, parallel branches, or stateful multi-turn interactions. The overhead is higher, but the flexibility is worth it for non-trivial workflows.

The adapter pattern means you can prototype with CrewAI's simplicity and migrate to LangGraph when complexity demands it -- without changing your agent definitions or tool integrations.