22,900 Stars: The 12-Factor Checklist Every Agent Team Is Saving
Humanlayer's 12-factor-agents reached 22.9k GitHub stars by naming what production teams already know: 80% quality with a framework is easy. The last 20% — customer-facing, on-call-worthy — requires owning your prompts, your context window, and your control flow.
NeuroX AI · May 31, 2026
12-factor-agents hit 22.9k GitHub stars — and the framing will feel familiar to anyone who shipped cloud apps in the 2010s. It's a production checklist for LLM systems, named after the methodology that defined cloud-native engineering. Dex at Humanlayer built it from one blunt observation: reaching 80% quality with an agent framework takes a week; the final 20% — production-grade, customer-facing — typically requires abandoning the framework entirely.
The 12 factors are engineering decisions, not model decisions. Three explain most prototype stalls:
- Factor 2 — Own your prompts. Frameworks abstract prompt management. Production systems need deterministic control — you need to know exactly what the model sees on every request.
- Factor 3 — Own your context window. Frameworks fill context by default. Production agents require surgical curation: every token costs latency and money at scale.
- Factor 12 — Stateless reducer. Agent state inside a framework is state you cannot inspect, replay, or debug when something breaks at 2am.
The repo's title sounds aspirational. The content is diagnostic. Every team that has shipped past week two in production has already hit each of these factors — usually the hard way, usually on a customer's timeline.
The bottleneck isn't model capability. It never was. The bottleneck is the engineering discipline the model doesn't bring with it.