1
Curious builder0 XP earned · 300 to level 2
0 daysFinish a lesson to begin
Badge collection0 of 6 unlocked
51 small wins to finish your pathNext question →
How would you choose between LangGraph, CrewAI, OpenAI Agents SDK, Google ADK, or writing your own loop?
30-second answerSay your answer out loud first, then reveal.
Evaluation criteria
- Control vs abstraction: can you see and modify exactly what's sent to the model?
- Production features: persistence/checkpointing, human-in-the-loop interrupts, streaming, retries, tracing.
- Model flexibility: vendor lock-in vs multi-provider.
- Multi-agent support: handoffs, supervisors, sub-agents.
- Ecosystem: integrations, MCP support, community, docs, maturity.
- Debuggability and testing.
Rough positioning (as of 2026; frameworks evolve fast)
| Option | Sweet spot |
|---|---|
| Raw loop (~50 lines) | Simple agents, maximum control, learning |
| LangGraph | Stateful, complex control flow, durable execution, HITL |
| CrewAI | Fast prototyping of role-based crews |
| OpenAI Agents SDK | Lightweight agents with handoffs, guardrails, tracing |
| Claude Agent SDK | Agents with a Claude Code-style harness (files, shell, sub-agents) |
| Google ADK | Gemini / Google Cloud ecosystems, multi-agent |
| Pydantic AI | Type-safe agents, structured outputs |
Strong answer. "The framework matters less than the fundamentals: tool design, context, evals. I'd pick the thinnest layer that gives me checkpointing, tracing and HITL, and make sure I can always inspect the raw prompts."
Common mistakes
- Picking a framework for its demo, then fighting its abstractions in production.
Related
Every expert started right here.