My AI fleet sent 1,126 emails in one day — here's the receipts

On July 10, between 11:22 AM and 1:41 PM, the agent fleet running on my laptop sent 1,126 emails across two campaigns. Not "scheduled" — sent, one API call at a time, with every send logged to disk. The plan for this post said the number was 1,167. I checked the logs before writing this, and the logs say 1,126. The real number is the headline. That's the whole point of this blog.

The two campaigns

Campaign one: a pottery video catalog we remastered to HD. 66 emails — 5 past buyers got a thank-you with a free upgrade link and no pitch, and 61 newsletter subscribers got the announcement with a coupon. Buyers were deduped out of the subscriber send so nobody got pitched on something they already owned. The whole run took under two minutes.

Campaign two, two hours later: a remaster announcement for a senior-fitness brand's 1,061-subscriber list — an old Mailchimp export, which is the riskiest kind of list you can mail. 1,060 sends were accepted; 1 API call failed with a gateway error (HTTP 522) and sits in the log as a FAIL line instead of being silently retried.

How you mail a 1,061-address export without torching your sender reputation

An old export is full of dead addresses, and mailing it naively is how domains end up on blocklists. So the send script is built like plumbing, not like a blaster:

• Test gate first. Both campaigns sent their drafts to my own inbox before any real recipient. I read them as a subscriber would, then approved. The test sends are in the same log, tagged -test.
• Engagement tiers. The big list was split by the export's own engagement rating. Tier 1 (engaged, rating 3–5) went first. Tier 2 was only allowed to run after tier 1's bounce numbers came back clean. Both cleared the same day — but tier 2 had to earn it.
• Throttle. One send every ~1.1 seconds. 1,061 sends took just over two hours on purpose.
• Idempotent resume. Every send appends a line to a log file, and the loader skips any address already marked OK. A crash mid-run can't double-send.
• Human approval on the words. The agents drafted, formatted, and sent. A human approved the copy and gave the go. Nothing goes to a stranger's inbox on model output alone.

The receipts, 24 hours later

Numbers from the email provider's aggregate report for the big campaign's tag, pulled by API the next day: 1,060 requests, 695 delivered, 4 hard bounces, 173 soft bounces, 199 blocked outright by the provider's suppression list, 8 deferred. Engagement so far: 118 unique opens, 43 unique clicks, 3 unsubscribes — and 0 spam reports.

Yes, that's a 65% delivery rate. That's what an aging export looks like, and it's exactly why the tiering and throttling existed: the 4 hard bounces and 199 pre-blocked addresses got absorbed without a single spam complaint. The 66-address pottery run, a clean opt-in list, went 66-for-66 with zero failures.

Same week, same pattern

The email day wasn't a one-off burst. The same approval-gate pattern published 3 Instagram reels for the pottery brand over July 10–11: the publisher agent hosts the video, creates the media container, and posts — but only after a logged human "APPROVED" line, and never more than 2 per day. The cap is enforced in code, not in good intentions, because automation that gets an account banned is worth less than no automation.

Why I'm showing you the log format

Every claim above traces to a flat file: a timestamp, a campaign tag, an address, and OK or FAIL. When I say 1,126, it's because grep -c says 1,126, not because it sounds impressive. The system that sent these emails is the same system that wrote the first draft of this post — and the rule it operates under is that nothing gets published, mailed, or claimed unless the log agrees.

Want this running on your hardware?

I'm taking the Founding 3 — a $299 done-for-you setup (then $499): I install the fleet on your machine, run the first-run interview, and get your first verified content live before I log off, delivered within 14 days. Two things before you buy: (1) do you have a machine with an NVIDIA GPU — 6 GB VRAM or more — that can stay on 24/7? (2) Is there one person on your team who'll spend 10 minutes a day approving what the agents propose? If both are a yes, join the waitlist and say so in your reply — you're in the Founding 3.