The best feature of Google's AI coworker is that you can fire it

Seven days. That's how long Google says its new Gemini agents can run on their own, no human at the keyboard. Call it 168 hours of unsupervised work, and most people are reading it as the start of the AI coworker era.

Everyone is reading this as an autonomy story. The short version: it is actually an accountability story. The headline feature isn't the week of solo work, it's the Gmail address.

Here's the thing. Google Cloud is giving each agent its own Workspace account, with email, a calendar, Drive storage and a slot in the company directory. When an agent edits a doc, its name shows up in the version history, not yours.

That sounds like an HR gimmick. It isn't. Every agent gets a cryptographically attested identity and least-privilege permissions approved by security admins, and every action is logged to the agent instead of a person.

Why does that matter so much? Because most companies can't tell who did what right now. CrowdStrike's CTO said at RSAC 2026 that most enterprise logging can't tell whether a browser was launched by a human or spawned by an agent in the background.

You can't govern a worker you can't see. And buyers are feeling it. The share of enterprises letting agents make production changes without human review fell from 75% to 56% between July and August.

Among final purchasing decision-makers, it dropped from 88% to 61%. That's a 27-point swing in one month, which is wild for a survey. The people who sign the checks blinked first.

The reason isn't a mystery. 61% of companies running pre-deployment evaluations had an agent pass internal testing and then cause a customer-facing failure, and only 5% fully trust their automated evals. The test said yes, and production said no.

Now stack Microsoft's warning on top. Satya Nadella wants an emergency brake built in, so an authorized person can pause or shut down a model mid-task. The same report says 60% of organizations can't terminate a misbehaving agent.

Let's be real here. If you can't stop it, you don't manage it, you just hope. An agent with a name, a permission list and a log is something you can actually stop.

This is where it gets fun. The risk isn't only in what engineering ships. Pluto Security says about 60% of the risks it detects come from AI tools built by non-developers, with finance generating the most exposure and developers ranking fourth.

So who wins? Vendors that treat agents like employees with badges, not scripts borrowing someone's password. Gartner backs the logic: it predicts 40% of companies will decommission agents by 2027 because they never separated what an agent can do from the access it was granted.

I'll be real, though. Google's version is a private preview with no general availability date, and Google hasn't detailed how data is handled when Claude models process enterprise content in the same workflow. A great design in preview is still a promise.

Who pays? Every team that already handed agents human credentials and a vague job. Harness found controls on testing, security, cost and rollback lag confidence by 30 to 55 percentage points, and that gap is the bill coming due.

My rule before any agent gets a week of autonomy: Can you find its name in the log? Can you see exactly what it's allowed to touch? Can someone shut it off mid-task in under a minute? Three yeses, let it run. Any no, keep a human on review, like the 42% of enterprises now planning to do exactly that.

See you next week, geeks.