Five controls that keep LangGraph agents alive in production: 15% of chats routed to humans
LangChain · x · 2026-11-02
A LangChain Community Spotlight by Simon Budziak distills five controls behind two production LangGraph systems (an AI SDR for Open Loyalty and a patient education assistant for Shen.ai):
- Model the agent as an explicit state machine;
- Trace every production turn;
- Put an evaluation gate on every change;
- Make every side effect idempotent;
- Route low confidence to a person in code — in the patient assistant, a confidence gate sends 15% of conversations to a clinician, and clinicians judged 80% of those handoffs appropriate.
Citing LangChain's survey of 1,340 respondents (57.3% already run agents in production; quality is the top blocker at 32%, ahead of latency at 20%), the thesis: what kills agents in production isn't capability but the surrounding system failing to notice when the model is wrong. Durable checkpoints underpin it all; code shapes are included.
More from coding & agent
- DeepSeek Harness Super-agent design: the bottleneck isn't smarter models, it's your attention — sujingshen · 2026-11-02
- tito: the hidden tokenizer bugs breaking RL training, and a Lean-verified fix — VoidAsuka · 2026-11-02
- Build an Agentic Knowledge Graph from Scratch — No Graph DB, No Extraction Model — adnan_hashmi · 2026-11-02
- One Vault, Two Sides: An Obsidian + Claude Agent Workflow for Personal Knowledge — adnan_hashmi · 2026-11-02
- Vercel engineer open-sources Security Audit skill you trigger via @bot reply — brandon_galang · 2026-11-02
- Vercel Labs shipped 95 releases in October, debuts agent-desktop for computer use — evilrabbit_ · 2026-11-02