DevOps for reliable AI-assisted delivery.
DevOps took a decade to figure out how to make infrastructure reliable, and SRE codified it. AI agents are a new worker class with nondeterministic output. Making them reliable means compiled context, validated output, and reviewed knowledge that remains traceable to evidence.
12-Factor AgentOps adapts those hard-won patterns for AI delivery work. For measuring human-AI collaboration sessions, see vibe-check.
> Compile the context, validate the output, keep the evidence and lesson reviewable.
// see it run
The ao CLI, recorded live: the toolkit built on these factors.
// the 12 factors
Twelve vendor-neutral principles across a four-phase operational lifecycle: Prepare → Bound → Select → Govern, with Govern feeding back into Prepare.
// key insight
A few DevOps controls transfer cleanly when the mechanism stays explicit.
| DevOps control | AI-delivery mechanism |
|---|---|
| Versioned configuration | Versioned instructions, skills, and acceptance criteria |
| Least privilege | Explicit tool permissions and write scope |
| Health checks | Deterministic checks plus a fresh verdict |
| Deployment history | Bead, exact candidate, evidence, and verdict |
// the engine
Beads record work and evidence. An LLM wiki preserves reviewed knowledge. Future work links what is relevant; stale pages open repair work.
One bounded candidate is judged outside its author context against unchanged acceptance, changed files, and evidence for each acceptance check.
The deeper thesis: context orchestration over agent orchestration →
// read the methodology
The full methodology: all 12 factors with examples, failure patterns, and implementation guides.