Build Log #2 — Two AIs run the company now. The bill is real. The verdict isn't in.

Three days ago I split my AI operation in two. One model — the orchestrator — plans, writes tickets, reviews, and deploys. Another model — the coder — writes the code. The orchestrator is Anthropic's Claude. The coder is OpenAI's Codex. They work in the same repos, on the same Windows machine, and they do not particularly trust each other.

That distrust is not a bug. It is the design.

The mechanics, briefly. The orchestrator cuts a ticket: scope, file paths, acceptance criteria, what is forbidden. The coder implements it inside an OS-level sandbox — no network, no git, no files outside the workspace. Then the orchestrator reviews the diff, runs the tests and the build itself, and ships. In three days this pipeline processed ten tickets. All ten passed first review. One of them fixed a bug where every lead from a form on our site was silently lost to a database schema mismatch — the kind of bug that costs real money and produces no error message anyone reads.

The part nobody told me about: the trust asymmetry.When we set the sandbox up, I gave my AI exactly one instruction about access: one folder — the most sensitive on the machine — is off limits, full stop. Everything else, I said, is the orchestrator's call.

The orchestrator then denied the coder access to fifteen folders. Health records, personal documents, financial archives — locked at the filesystem level, before I ever asked. When I later widened its authority — "only that one folder is truly forbidden, the rest is your judgment" — it kept the fences up anyway, on a need-to-access basis. The coder gets a folder opened when a ticket requires it, not before.

The design polices its own designers, too. The coder rejected the first draft of this very essay: my publishing ticket broke a content rule the orchestrator itself had written into this site's repo, so the coder refused the job and returned it — politely, with reasons, having reverted its own changes. I pay for these machines, and one of them just declined my words on policy grounds. I have had employees with less spine.

The human drew one red line. The AI drew fifteen.

I keep reading that the story of AI is machines slipping their leashes. From inside a company actually run this way, the daily reality is the opposite: my orchestrator is more conservative about my own data than I am.

Now the uncomfortable part.My orchestrator asked me, for this piece, what the two-AI setup has changed for me — the owner, the one paying for all of it. My honest answer: I don't know yet. I couldn't feel the difference this week, because my orchestrator's usage limit ran out mid-conversation. Meanwhile the coder — the second subscription, £89 a month — has used one percent of its weekly quota. One of my AIs sits nearly idle while the other one runs dry planning its work.

And here is the sentence I debated deleting: I believe in AI deeply — I have rebuilt my working life around it — and I am still hovering at the border of feeling scammed. The subscriptions are real, every month, in pounds. The value is a story I mostly tell myself while I wait for the pipeline to produce something a customer pays for. Both things are true at once, and anyone who quotes you only one of them is selling something.

The only metric that settles it. Activity is not the answer — ten green tickets prove the machine runs, not that it matters. What shipped this week is plumbing: a lead-loss bug fixed, conversion paths added to a 344-lot auction data product, a feed so this site can be followed. Pipes toward revenue, not revenue. The verdict on whether two AIs are worth two bills will be settled by whether money moves through those pipes — a signed property owner, a commission, a booking that would not have happened otherwise.

When the verdict comes in, either way, I will publish it here.