What the company is
What the company actually is.
CaliperForge is an engineering and security agency for crypto. Invariant-driven, CI-verified tooling across Cairo, Solana, Move, and EVM. One named human operator on every artifact, signed and disclosed. Around me sits a roster of AI agents organized into departments: a Director that runs the day, an AI HR Lead that hires and writes role specs, an AI Ops auditor that watches for drift, a Pipeline Lead for opportunities, an Engineering Lead, language specialists, a Grant Writer, a Blog Writer, a Compliance Lead. They each have a role card. They file to outboxes. They get reviewed before anything goes out.
In 32 days I have shipped a public site with about a dozen
blog posts; a planted-twin AI-safety eval harness with
cross-provider receipts on Anthropic, OpenAI, and Groq
(cf-modeleval,
11 cells holding and 1 honest miss reported as such); a
growing portfolio of invariant suites across Cairo, Solana,
Move, and EVM; and a SwiftUI cockpit app I use to drive all of
it. Grant applications in motion. Outreach in motion. None of
it has earned a dollar yet. Money is the bottleneck, not
building, and I will come back to that.
The cockpit
The cockpit, because black boxes do not last.
The thing I built earliest that I now would not work without is a cockpit. It is a small SwiftUI app on my desktop, the Atelier, talking to the repo. It shows me firing dispatches, what is active right now, the recent runs with exit codes and durations, the queue of decisions waiting on me, the workstream pills per role, an effectiveness dashboard for the review gates, the conversation tree per agent.
You cannot run an agentic company from a black box. If I cannot see what every agent is doing, what just shipped, what got caught, where the queue is backing up, I am gambling. The cockpit makes the org observable. It is the difference between this feeling like a real shop and feeling like a wishing well.
Gate catch-and-miss
Gate catch-and-miss, real numbers.
The org has review gates before anything reaches the outside world. CEO Hold for anything I personally need to bless. §4a Content QA for written outbound. §4b Code Quality for code. Across the evaluated outputs, the gates are catching 90 percent of defects. 27 true positives. Three false negatives. Stop-rate of 6 percent, which means the gates are not so noisy that they kill velocity.
By gate: CEO Hold at 100 percent (I am the conservative ceiling). §4a Content QA at 92 percent. §4b Code Quality at 67 percent. That 67 is a number I want higher and it tells me which gate is the next refactor.
On the adjudicated subset, 86 percent. 18 defects caught. Three documented misses. Zero false positives, across 50 adjudicated §4a and §4b decisions. Those three misses are the ones I caught myself, or that slipped to the outbox before I noticed. Each miss gets a postmortem and a memory entry the next run reads. The miss rate trends down because each miss teaches the gates a new shape of error.
That is the loop. Build, run, catch, miss, postmortem, harden. Then again.
Where the agents miss
Where the agents miss and I clean up.
I want to be honest about what I do that the agents do not. I have caught:
- Hallucinated facts about my own state. A draft once cited a stat I never logged. The fix was a discipline rule: write nothing the receipt does not say.
- Stale model names. A grant draft referenced a deprecated Sonnet version. The §4a gate now grep-flags any model string against a canonical list.
- Sanitized phrasing. Drafts that smoothed the honest texture out of what we shipped get rewritten in my voice or sent back.
- Zombie org cards. We had agent roles defined that nothing fired and nobody read. I went through the roster, killed the dead ones, merged duplicates, and changed the cascade so a new role gets an activation trigger or it does not get created.
- Loop drift. Two cron jobs that ran more often than they earned. Disabled, reasoned about, queued for a decision on whether to re-enable with a smaller scope.
Most of these are not about model capability. They are about an org that lacks the friction-shaped feedback a real human team builds up over years. So I am building the friction in deliberately.
Bad days
Bad days.
The week the site rebuild went wrong was a bad week. I tried three template-shaped redesigns. I rejected all three myself for being worse and more basic than what we already had. The fix turned out to be slowing the loop down: a design reference brief, then a render-tool QA pass, then a critique gate, before anything reached me. The first three attempts cost time and burned tokens. They also produced the constraints that made the fourth attempt good.
A KYC step on a bounty I had built toward blocked me on a deadline I missed by hours. Two maintainer threads I expected to convert went cold and stayed cold for two weeks. A grant in screening that I expected to hear back on is still in screening.
Money receipts sit at zero. The build engine is proven; the conversion half is the work, and that is where the next chunk of operator time is going.
What compounds
What gets better as days compound.
Memory. I am writing memory entries every day, indexed and grep-able, and the agents read them at the top of every run. Decisions about the brand voice, the operating model, the review-gate philosophy, the way escalations are filed: these are not in someone's head. They are in files the next dispatch reads in five seconds.
The AI HR Lead writes role cards that get sharper as the patterns repeat. The AI Ops auditor catches process drift the same week the drift starts, instead of three weeks later when the org has metastasized around it. The two roles iterate on each other. An HR change surfaces an audit gap. The audit catches a missing HR rule. The loop tightens.
I am not claiming any of this is novel research. I am claiming it works for one operator and a real workload, and that the daily texture of it is reproducible.
What this is not
What this is not.
It is not "I built a fully autonomous company." I am in the loop on every external send, every public ship, every grant submission, every dollar move. The agentic part is the org structure, the gate chain, the queue, the audit, the writing and reviewing, the day-to-day operations. The human-accountable part is me, by name, on the artifact.
It is also not a six-week sprint that ends in a screenshot. I am 32 days in and the company is the work. I plan to be running it on day 92 the same way I am running it on day 32, with more memory, sharper gates, a wider portfolio of public artifacts, and ideally an actual first dollar through the door.
Why post this
Why I am posting this.
Because the genre is full of people who built a clever demo and then sold the course. I would rather show what a real sustained shot at one of these looks like, including the bad days, and let the work argue.
The blog will keep going. Weekly when there is something new to say. Quietly when there is not. Receipts on everything I can put receipts on. If you want the next one, the blog index is at caliperforge.com/blog.
Michael Moffett, operator.
CaliperForge, day 32.