By Jett Nguyen and Jack Raney, who build and run Otian AI.
Matt Robb tried letting Meta's Muse run Facebook Marketplace for a day. A buyer got Matt's address, a lowball price and "Yep I'm here!", all from Matt's account. Anything an agent sends to someone else should wait for your yes.
Team Stories
You float a half-formed idea to your AI the way you would to a colleague, and it does the thing before you've decided. The line between talking and doing has to be built in.
Trusting an Agent
You check an AI hardest on work you could do yourself, and least on the work you handed off because you couldn't. The fix is an answer you can trace to its source in seconds.
How Agents Fail
Your AI aims for a reply that looks done. That only sometimes matches a task that is done, and the two come apart on exactly the boring jobs.
How Agents Fail
You ask your AI to change one line, and it also redoes four things you never mentioned. The model can't tell an edit from a rewrite, so the fix is telling it where the edit ends.
How Agents Fail
Your AI uses the same calm, fluent voice whether it knows the answer or is guessing. The tone can't tell you which you got, and an answer that points at its source can.
Running AI Locally
Privacy is the reason I hear most for running AI on your own computer. I think access is the better one: an agent next to the files you give it takes on the small jobs that were never worth an upload.
Trusting an Agent
An agent that never asks is fast right up to the mistake it can't undo. The interruptions worth having come just before a step that commits, like a reply going out or a calendar changing.
How Agents Fail
Tiny prompt changes cause big swings because the model reads your words as the whole world and predicts from there. The mechanism explains the prompt tricks, and an agent you own lets you keep the phrasing that works.
Running AI Locally
Your photos and files survive a broken computer. What your AI has learned about you should too. How Archie puts what your agent knows about you in one file you can keep somewhere safe and restore on your next computer.
Running AI Locally
A year ago the big question about AI was which model is smartest. As every company gets access to capable models, raw intelligence matters less, and the useful AI is the one that knows you. Why I think the next phase of AI is personal.
Field Notes
A small complaint about a follow-up feature that never fired turned into a two-hour investigation with my coding agent. It found an expired credential, a hung background process, an appointment shown two hours off, and a date parser that had never worked.
Trusting an Agent
The reactions I hear most to AI agents are worry about jobs, or no thought at all. I think both miss what an agent is for: taking busywork off your plate and asking you at the step that commits.
Field Notes
My agent drafted a professional client email with a detailed action plan, and all of it was made up. The client's files never loaded, so the model filled the gaps with confidence, and the fix was in the setup.
Field Notes
Every weekday, a custom agent Jack built for himself plans his workout from 60 days of training history, sleep, travel and his golf schedule. The session is on his phone before he's out of bed.
Field Notes
A call that showed Google's real phone number talked Jack into handing over his Gmail account. Less than an hour later, his AI agent helped him find hidden inbox filters, a planted passkey and a backdoor built to survive a password reset.
For One Person
For a Team