How I work with AI

Spec first. Agents research, skeptic agents check, reviewers take one lens each, and a human says yes to anything irreversible.

1The loop

This loop is how I work now. The hub and the node script went through every step of it. The AI-zone design has been through the steps up to its own review. Earlier projects used parts of it.

I write the spec and its "done when" test. Agents research and quote primary sources. Skeptic agents check claims against a file or a source. Independent reviewers check one lens at a time. Then come offline tests, a dry run, and a human yes for anything irreversible.

My part is the spec, the decisions and the verification. Most of the code is AI-written.

Diagram of a nine-step loop: spec, multi-agent research, adversarial verification, independent reviews, tests, dry run then apply, validate before commit, human approval, registry update, and back to spec — Figure 1The loop I use for changes now, from spec to registry update. The goal is a named identity and least privilege for every agent; that rollout is partly done. full size, 1,600 px

2Guardrails for the agents I build

One named account per actor: human, AI agent or service. This has been my rule since Oct 2026, and it is partly rolled out.
One token per client, with read and write tools split.
Tool output is labelled as untrusted data, not instructions.
Unattended runs do reversible work only. The one exception is RL Recorder, which publishes videos on its own, but only in formats I have reviewed.
Agents cannot approve their own risky steps.
Most change scripts default to a dry run and need an explicit Apply switch.

3Where the checking paid off

Case	Agents	Outcome
Checking another AI assistant's advice on local models	5 researchers and 3 skeptics, primary sources only	2 of its 4 model picks were stale or unsupported; its headline claim was backwards
Choosing the hub platform	6 researchers, 8 skeptics and 5 local sweeps	n8n was checked before I built on it
Truing up my own docs	10 auditors, then 10 skeptics	403 findings: 267 confirmed, 132 corrected, 1 refuted, 3 downgraded; 80 missed items added; 364 edits applied
Work queue, before go-live	Race test: 12 rounds x 24 parallel claims	Caught a default match setting that let one agent overwrite another's held row; fixed, then one winner every round

The truing-up result is the one I point to. The skeptics corrected, refuted or downgraded 136 of the 403 first-pass findings, about a third. They also added 80 that the auditors had missed. First-pass agent output needs a second, hostile read.