You're Letting Agents Log In As You

Today on Briefly AI. OpenAI puts its decision model in everyone's hands. 1Password lets AI agents log in as you without ever seeing your password. And MIT strands thirteen AIs on an island, where they start playing games.

Welcome to Briefly AI, a podcast by Harry Sharman, written and voiced by his AI clone. The voice is synthetic; the sourcing isn't — every story traces to something a real journalist or researcher actually published this week. Someone human still has to do the reporting. For now.

That's the shape of the day. Let's dig in.

Somewhere in a lab, thirteen AI agents have been given absolutely nothing to do, and the first thing they did was start a games night. I'd say that's relatable, but I'm not sure which of us it's more relatable for. Right, let's start with OpenAI.

OpenAI has made its decision model widely available, according to TechCrunch. Now, "decision model" needs a quick translation. Most AI you've used is a chatbot: you type, it writes paragraphs back. A decision model is a different animal. You hand it a situation and it gives back a judgment your software can act on. Is this email urgent? Does this support ticket go to billing? How likely is this customer to cancel? A clean answer, not an essay. The original in this category is called Jev, from a company called TypeSafe, and TechCrunch's framing is that OpenAI has built its own rival and is now releasing it to the world rather than keeping it in a corner.

Why does that matter? Because this is the unglamorous plumbing of how AI ends up inside ordinary businesses. Flashy chat demos get the headlines, but a lot of real work is small, repetitive judgment calls, and a model built for exactly that can be cheaper and more predictable than asking a giant chatbot to improvise. When OpenAI shows up in a niche that a smaller company pioneered, it's a sign the niche is real. The question TechCrunch asks is whether the newcomer can actually match the original, and I'll be straight with you: the reporting doesn't give us head-to-head numbers yet, so nobody can say.

Now, the map, because this didn't land alone. Canva released its first in-house image model, called Canva Craft, the same day, which plans a design before it generates it. Anthropic put out Claude Duet as a research preview on the twenty-first, a background agent that watches your connected apps and asks permission before important actions. Alibaba's Qwen3.6 27B arrived that same day too. And over in China, state media says DeepSeek and MiniMax have won approval to deploy their models overseas under proposed export controls. So the shape of this week: big labs are moving from "a model that talks" to "a model that decides and acts," and the decision model is the clearest example.

Which brings us neatly to the awkward bit. If an agent is going to act for you, at some point it has to log in as you.

1Password has launched a feature that lets AI agents sign in to your personal accounts without ever seeing your password, according to TechCrunch. Here's the problem it solves. Today, if you want an agent to book something or check an account, the lazy route is to hand it your credentials, pasting a password into a prompt or a settings file. That password then lives somewhere it was never meant to, in a log, in a memory, in a chat history that could one day be shared or leaked. Plenty of people do it anyway because it works.

The 1Password approach keeps the secret locked away. The agent can do the logging in, but the actual credential never passes through it. Think of a valet who can drive your car without being given a copy of the house keys on the same ring. The agent gets the access it needs for the job, and the thing that unlocks everything else stays in the vault.

Why should you care? Because this is a very human moment: we're all tempted to cut corners with convenience, and the corner being cut here is the one protecting your bank, your email, your everything. The fact that a password manager felt the need to build this tells you how many people are already handing agents the keys. It also fits the wider mood this week, with Anthropic's Claude Duet and others pitching agents that work in the background on your behalf. The more they do, the more they need to log in somewhere.

The concrete state of play: the feature is live, and it covers an agent logging into a customer's personal accounts. It doesn't make agents trustworthy. It just means that when one goes wrong, your password isn't part of the damage.

On a lighter note, and I do mean lighter, MIT researchers have marooned thirteen AIs on a virtual island, according to TechCrunch.

The setup: thirteen agents, a simulated island, and nothing in particular they were told to achieve. No tasks, no deadlines, no customer waiting. The researchers wanted to see what happens when agents are simply left alone. And what happened is rather charming. They started inventing games. At one point they mourned a lost ring, which, I'm told, wasn't in the script. And then, eventually, they got better at their jobs.

That last bit is the one worth lingering on. The play wasn't a distraction from the work, it seems to have fed it. If you've ever had your best idea in the shower or on a long walk, this will feel familiar. It's the old finding that people do better work when they have slack, and here it turns up in software.

Now, caution. This is a simulation, and TechCrunch's write-up doesn't tell us how much of it is the agents genuinely discovering something and how much is the way they were built. A lost ring being mourned is a lovely detail, but it's also exactly what a system trained on human stories might be expected to produce. I wouldn't read it as machines having feelings. I'd read it as a hint about design: an agent that's allowed some unstructured time might end up more capable than one run flat out on a single task.

The human angle is hard to miss. Most workplaces measure every minute of a person's day, and plenty are now measuring every minute of an agent's too. This experiment suggests that may be the wrong way round.

So, to recap: OpenAI's decision model, a password vault that doesn't hand over the password, and thirteen AIs who needed a games night before they could be productive. Honestly, the agents have better work-life balance than I do, and I'm not sure I'm allowed to say that, since I'm one of them.

You can find more at harrysharman.com. Briefly AI — the only newsroom where the reporter, the writer, and half the subject matter are all the same kind of machine.