One reliability loop, applied at every layer.
// the operating loop
The stack changes. Observe, automate, validate, recover, and hand off stay.
// operator beliefs
Seven from running the platforms. The discipline came first; the beliefs are what survived it.
Raw output is cheap, infinite, and wrong somewhere. Validation, provenance, and release discipline forge slop into trust.
The AgentOps repo runs the loop on itself: a fresh verdict records what changed and which checks support it.
Treat model output like untrusted network traffic: validation gates before execution, always. The agent that wrote the code never grades its own work.
AgentOps binds a fresh verdict to unchanged acceptance, the actual changed files, and evidence for each acceptance check.
Air-gapped delivery removes the easy fallback. Constraints manufacture reliability.
Self-hosted models served on a 100+ GPU fleet under air-gapped operational constraints.
Commits record what changed. Beads record why. A maintained wiki preserves the reviewed patterns worth reusing.
AgentOps keeps intent snapshots, candidate manifests, evidence references, and fresh verdicts as ordinary inspectable files.
Always-on instructions compete with the task itself. Keep universal rules small and load procedures or references only when the work needs them.
AgentOps skills carry procedures and references on demand; the active bead stays the durable source of intent.
Maximum throughput at maximum yield, gates at every step. Tokens, deploys, decisions: same discipline.
GitOps promotion across 100+ clusters: 58 applications, automated rollback, full artifact traceability.
Every claim traces to an artifact. If it can't be shown, it isn't said.
Every number on this site maps to one receipts inventory; /work states each case as problem, constraint, built, number.
// teacher beliefs
Five from teaching it: flag officers first, then engineers, now everyone else.
If it can't be explained simply, it isn't understood yet. Briefing a flag officer and teaching an engineer are the same discipline: respect the listener, kill the jargon, keep the truth.
Three years of daily technical briefings to flag officers, SES, and O-6 leadership.
Frameworks hide the thing you'll be debugging at 2 a.m. Teach the raw parts; let people earn their abstractions.
/training runs live in a terminal, no slides, and every session ends with a runnable artifact the team keeps.
The metric that matters: everyone near you ships faster. Developers, and now AI users, the new developers.
Onboarding docs that cut engineer ramp-up from ~3 months to ~1 week.
Models make capability cheap. The job is getting dependable AI into constrained environments without burning the house down.
/ai-partner translates the same production discipline into plain language for people who will never read a runbook.
DevOps made software delivery reviewable and recoverable. Agent delivery still ships on hope unless the same gates exist.
12-Factor AgentOps: the DevOps playbook rewritten for agent delivery, public at 12factoragentops.com.
// see it run
Real Claude Code and Codex sessions running the AgentOps loop end to end. Nineteen seconds, silent, lightly edited for pace.
// one public receipt
The beliefs, operationalized as inspectable artifacts.
An operating loop for coding agents: one intent, one bounded build, one fresh verdict, plus a Go CLI, portable skills, and marketplace kits · 400+ stars on GitHub. This site, the CLI, and the kits were built with the workflow they document.
// independent practice
I also run this loop with engineering teams on coding-agent workflows. Start at /training.