Three production LangGraph agents converged on the same five controls for keeping agents alive
hwchase17 · x · 2026-10-25
Simon Budziak (retweeted by LangChain CEO Harrison Chase): getting a LangGraph agent working takes an afternoon; keeping one in production is the actual job.
Across three shipped projects—an AI SDR, a patient education assistant, and an AI marketing localisation pipeline—they converged on the same five controls:
- An explicit state machine instead of free-form dialogue flow
- A trace for every turn
- An evaluation gate on every change
- Idempotent side effects
- Confidence routing to a human
The full breakdown, with the code shape behind each control, is in the follow-up replies.
More from coding & agent
- Creator builds year-long video pipeline pairing Hailuo H3 with Magnific MCP — TinfoilTricorn · 2026-10-25
- Why stuffing 10+ AI bots into a group chat makes outputs worse — gaganghotra_ · 2026-10-25
- Biotech found 30% of in vivo studies redundant — agent prompt to review studies pre-execution — nlarusstone · 2026-10-25
- Reward design details — latency penalties vs collaboration bonuses — decide multi-agent RL — rosstaylor90 · 2026-10-25
- macOS trust bundle still trusts certs forged with leaked eMudhra keys; Codex CLI fixed in 0.164.0 — ai · 2026-10-25
- rakyll ships aic, a Docker-like container tool built for AI developers — rakyll · 2026-10-25