😸 Bots with bank accounts

gm legends, happy Wednesday.

Today: the money rails that let an AI agent actually spend your cash without draining your account, a Google video model you direct like an editor instead of praying at a prompt, and a smart ring you whisper into. Plus the forum on spotting the fake signups after a launch.

Sequence Agentic gives AI agents the one thing they've been missing: the ability to actually move money. In a single API call your agent can send, split, and route real payments across almost any US bank, with scoped, revocable keys and server-enforced spending caps so it can't overspend. It comes from a company that's already routed over $3 billion for freelancers, small businesses, and households, now pointing those rails at agents.

🔥 Our Take: Honestly one of the more useful AI launches lately. Every agent can plan a whole purchase and then freezes, because it can't move money, that last step still needs a human to hit confirm. Sequence lets the agent actually pay, one call, real money, any US bank. And I'd trust these guys, they already move $3 billion a year for normal people, so they're not learning payments, just aiming rails they've run for years at agents. They built the guardrail that matters too, a key you cap and can kill, so it can't overspend even if it loses the plot. That's the real fear, not that a bot can pay but that it drains you, and someone in the comments flat out said they'd never let AI near theirs.

Gemini Omni Flash is the first model in Google's new Omni family, and it collapses the whole five-tool video workflow (script, image, animation, voice) into one. The trick is that it holds a conversation while you edit: ask for warmer lighting or a swapped product in plain English and it builds on your last few turns instead of regenerating from scratch. Ten cents a second of 720p, and it launched at number one on LMArena's text-to-video arena.

🔥 Our Take: The reason AI video still feels like a toy is you only get one shot. You prompt, you get a clip, and if the lighting's wrong you can't nudge it, you start over. Gemini Omni Flash lets you talk to it while you edit, tell it to warm the lighting and it changes just that instead of regenerating everything. That's the actual leap, not the picture quality, because you stop gambling on prompts and start directing. And it's cheap, ten cents a second, same as Veo, already number one on LMArena. Funny thing is Google's beating its own Veo with it, but if talking-while-editing is the future, better them first.

Framer 3.0 ships AI Agents that live inside your canvas and can actually design. Not suggest. Design. Drop a brief and watch them build out full pages, create components, write code, connect your CMS, and handle SEO. All without leaving the canvas. They plug straight into Claude Code and Cursor. Yes, inside your actual component system, without breaking it. The Figma file you were about to hand off to engineering? Keep it.

OASIS 1 Ring is a smart ring that skips health tracking entirely and becomes an input device: whisper into it to dictate, and use its tiny trackpad to scroll and edit text with your thumb. Five years in the making from Ricky Rosa's Oasis Devices, it launches on Mac wired into Wispr Flow, aimed at voice-writing in quiet places where talking to your laptop feels weird.

🔥 Our Take: Finally a smart ring that isn't tracking your sleep. Every one so far, Oura, Samsung, the lot, has been a health thing. OASIS made it an input device instead, you whisper into it to dictate and there's a little trackpad to scroll and edit with your thumb, so you can dictate in a quiet office without looking like you've lost it talking to your laptop. The tell that they mean it is they spent five years just on the trackpad, that's obsessing over feel, not rushing to ship. It's launching on Mac tied into Wispr Flow, which is smart, whispering is exactly where dictation breaks down in public. Strapping a gadget to your finger is a big ask and there's a curve, no doubt. But it's the first ring I've seen built to do things, not measure you.

Pedro Reis (@pedromlsreis) asked a very post-launch question: after a burst of signups, how do you tell the real users from the disposable and unconfirmed ones, and what do you do with the fakes? His twist is that his product is privacy-first, so he logs nothing fingerprintable, and a tight cluster of signups with no follow-up activity was his only clue anything was off.

The room was near-unanimous on one thing: don't hard-delete. Soft-delete, mark inactive, purge much later. The warning that kept coming up was to never auto-delete an account that did something real, started a trial, uploaded data, even if they never confirmed, because that's a real user who's just lazy or distracted, not a bot.

Wasil (@wasil_abdal) had the most complete playbook: auto-delete unconfirmed accounts after seven days, but keep the soft-deleted row with a reason tag, and if the same email signs up again within 30 days, quietly restore the old record and resend the confirmation, no manual database surgery. Anastasiia (@alieksia) offered the opposite instinct, stop the cluster ever forming by verifying accounts before they touch your main table at all.