Avoiding LangGraph: Valid Reasons? | Generated by AI
Question: Why have I avoided LangGraph until now, and are my reasons valid?
Answer:
Your self-analysis is accurate. Let me validate and sharpen each point:
1. “My 200-300 lines + LLM calls already work” — This is the most honest reason, and it’s correct until it isn’t. You’ve hit the same wall most serious builders hit: hand-rolled agent loops are fine at 1-3 tools, 1 agent, linear flow. They start breaking down at:
- Multi-agent coordination (agent A waits on agent B’s output conditionally)
- Cycles with state accumulation (retry loops, reflection patterns)
- Checkpointing / resumability (long-running jobs that crash mid-way)
LangGraph’s actual value is the state graph with persistence — not the abstractions. If you’ve never needed checkpoint/resume or true conditional branching between agents, you genuinely didn’t need it.
2. Playground/demo directory disruption — Real friction. The cognitive cost of “where do I even put this experiment” kills exploration. This is fixable in one hour: create ~/code/langgraph-lab/ today with a barebones StateGraph skeleton.
3. No appropriately-sized project — This is the real gate. LangGraph makes most sense for:
- Multi-step pipelines where steps can fail independently
- Agent systems with human-in-the-loop checkpoints
- Your
ww/iclaw/zzCLI tools if they need branching retry logic
Your bank AI work is actually a perfect candidate — auditable, step-traceable agent workflows fit regulated environments well.
4. Knew about it since 2023 but waited — Classic “aware but not activated.” Right call honestly — LangGraph in 2023 was rough. The persistence/streaming story matured significantly in 2024-2025.
The real question now: What triggered you to use it seriously? That’s the signal worth tracking — what project forced the threshold?