More or Less · Topics · Episode #165

Brit's Agent Wall, Gmail's Missing API, and Why Nobody Makes Money

From #165 — AI Agents Are Going Rogue. Should We Be Worried? · September 4, 2026 · 5:37 min · Jessica Lessin, Dave Morin, Brit Morin, Sam Lessin

Clip from the full recording, 22:06–27:43 · download mp4 · watch at this point on YouTube

Brit now runs a dedicated agent display — Grokbot, OpenClaw, Codex, Gemini and Claude side by side. Sam goes deep on Gmail's missing draft-vs-send permission, then argues that data sovereignty makes every model interchangeable — which is great for users and terrible for anyone hoping to build a moat.

Bot check-in. Brit used Fable 5.1 to spawn subagents and it “didn’t go rogue, it completed the job really well”; her real week-over-week upgrade is hardware — a second display that’s purely an agent wall: Grokbot, OpenClaw, her terminal, Codex, Gemini, and Claude running side by side, racing on the same tasks to see who’s faster. Her one hard rule: bots can draft email but never send, with human edit on everything. That trips Sam’s wire: he just went deep on this and discovered Gmail literally has no draft-without-send permission — the compose scope covers both — so you can tell your bot to never send but you can’t enforce it at the API level. “You’d think they would decouple those now.”

Then the economics. Sam is a “split brain”: as a user, his token usage is inflecting up every week and it keeps getting more fun; as an investor, he can’t find the margin. His argument is structural: because everything is scriptable, on/off API switches don’t work anymore, and as long as you have data sovereignty — you own your databases, not the model — every model is one credential away from being a perfect substitute. Claude refused a task; Grokbot got the same login one second later. “They’re all completely interoperable and there’s nothing they can do about it, which is why I don’t understand how anyone’s ever gonna make money on any of this.” The winners he can name: Nvidia and DigitalOcean, because his data getting more valuable makes him buy more compute and storage. Jessica pushes on whether the sovereignty is real — Josh Wolf and others say the labs are training on everything — and Sam concedes instantly: “they’re definitely stealing all my data as we go.” But he holds the position anyway: the data at rest is his, synced to a hard drive in his office. “What are you gonna do?”

Key points

• Brit's setup: a dedicated agent display running Grokbot, OpenClaw, Codex, Gemini, and Claude simultaneously, racing them on identical tasks; Fable 5.1's subagents completed cleanly.
• Her guardrail: bots draft, never send, human edits everything — which turns out to be unenforceable at the API level.
• Sam's discovery: Gmail has no draft-only permission. Draft and send are the same fundamental API scope, so 'never send' is a promise, not a control.
• Sam's interoperability thesis: everything is scriptable now, so as long as you own your data, every model is a one-second credential swap away from being substituted.
• The margin problem: token usage inflecting up, more fun every week — and no visible moat for anyone but Nvidia and the infrastructure layer (his pick: DigitalOcean).
• The sovereignty caveat, conceded to Jess: the labs "are definitely stealing all my data as we go" — but the data at rest is his, down to a hard drive in his office.

Where they landed

Quotes

“As long as you have data sovereignty, they're all completely interoperable and there's nothing they can do about it — which is why I don't understand how anyone's ever gonna make money on any of this stuff.”— Sam

“I really wanted to give it draft but not send permission. And it literally doesn't exist in the Gmail stack.”— Sam

“Oh no, to be clear, they're definitely stealing all my data as we go. I'm just saying — I own the hard drive. What are you gonna do?”— Sam

Suggested tweets

All topics · Full episode #165 + transcript · Subscribe on YouTube · Spotify