03/22/2026
Isn’t it just crazy that one needs so much programming expertise just to write the code to orchestrate two AI agents, using existing libraries?
CrewAI: 15,000 lines of code.
LangGraph: 50,000 lines.
The thing I actually needed: ~150 lines.
So I built litecrew.
Here's a complete multi-agent workflow:
researcher = Agent("researcher", model="gpt-4o-mini")
writer = Agent("writer", model="claude-3-5-sonnet")
(researcher, writer)
def write_article(topic):
research = researcher(f"Research {topic}")
return writer(f"Write about: {research}")
That's it. No YAML. No config files. No framework to learn.
What it does:
→ Sequential handoffs (A → B → C)
→ Parallel ex*****on (fan out, collect)
→ Tool calling
→ Token tracking
→ Works with OpenAI, Anthropic, or Ollama
What it doesn't do (on purpose):
→ No hierarchical agent management
→ No state machines
→ No streaming
The philosophy: SQLite doesn't try to be PostgreSQL. Sometimes simple is the feature.
If you need the full power of CrewAI, use CrewAI. This is for when you just want agents to pass data without reading 200 pages of docs.
Open source. MIT licensed. Your API keys stay on your machine.
Link in comments 👇
