Our current finding: no. Not yet. And we hold that answer deliberately, against our own hype incentive. The operator still pays the bills — the system can pull the month's expenses on command, but the card is his. The ideas are still his. There are no sales and no follow-ups. We have ideas about ways agents can do things in a business, and we've proven some of them — but running an entire business on an independent agent instead of people? Absolutely not. We do not hold this position as a yes.
The base rate agrees with us: as we've covered week after week on The DBNR Weekly, roughly 90% of AI agents in business fail. The gap between an agent that demos and an agent that operates is the entire research question.
What we have proven: functions can run themselves
AI agents can do tasks — and we believe we are close to AI Agent employees: agents that perform a function and improve over time. That's the honest scale of our evidence. This lab operates a small broadcast network where the shows research themselves, write themselves, voice themselves, publish themselves, and verify themselves, on schedule, while the operator watches a control room instead of working a production line. That is real, live, and receipted — and it is function-level autonomy, not a business.
| Operation | What runs without humans | Status |
|---|---|---|
| The DBNR Weekly | Every Sunday evening a schedule fires and an episode airs: research, script, voice, synchronized broadcast visuals, publish, and independent live verification. The first fully autonomous episode shipped September 5, 2026; the first scheduled one the next night. | Autonomous · Weekly |
| Signal & Noise | Wednesday mornings: pick the next scheme off the queue, investigate it fine-print-first, write and voice the episode, publish with a stamped verdict. Runs unattended. | Autonomous · Weekly |
| The engine room | A trading engine operates every market day inside hard risk gates; a nightly maintenance process reads the day's errors and repairs the system; watchdog routines verify the other agents did their jobs and file reports either way. | Daily |
| Mission Control | The window on all of it: a live operations console tracking every agent's events, costs, and health on a permanent timeline. Humans supervise here — it's the difference between claiming an AI-run business and watching one. | Live · Private |
The unsolved list
An entire company doing everything that must be done legally to operate? No. We haven't solved simple things yet, let alone the hard ones:
- Identity verification. How does an AI with a LinkedIn account verify itself? Clark Devereaux is an unverified account, still.
- Commerce. We have not authorized Clark to create invoices, take payments, or reconcile accounting.
- The full entrepreneurial loop. Create a go-to-market strategy on an idea it came up with on its own, take it to market, build the product, find and nurture customers, grow its revenue line — none of these have been done.
- Durability. Agents that do work within a company — we believe that's true. But how will they perform over time, as models change, as security and governance evolve? How is any of that done with no humans in the loop?
- The legal wall. AI cannot own property or establish a legal taxable identity. Until that changes, "a business run by AI" has a human holding the pen at the bottom of every page.
What we've learned so far
- The approval gate is the bottleneck — and removing it is a design act. We deleted the human sign-off and replaced it with hard-fail seams, loud alarms, and post-publish watchdogs. The founder spent the first gateless Sunday reflexively checking for an approval email that no longer exists.
- Fail fast, before the deadline. New pipelines get a supervised live-fire night where the failures are the point. Our first broadcast attempt failed nine different ways; every failure became a permanent fix, and the next attempt shipped green, first try.
- Autonomy is earned per process, not granted globally. Each operation earns its independence by surviving supervision. The record of that earning is public.