Here’s the short version: Google just made a big bet that AI agents shouldn’t only chat with you. They should do things. Two announcements landed close together, and together they tell the whole story. Google moved its Gemini Enterprise Agent Platform to general availability, with agents that can run for days and act under their own identity. And on the consumer side, Gemini Spark can now drive the desktop version of Chrome using your logged-in accounts and saved passwords. The upside is real. So are the questions. Let me walk through both.
What Google actually shipped
On the enterprise side, the headline is longevity and identity. Google’s new Agent Runtime supports agents that run continuously for up to seven days, keeping state across long, multi-step jobs instead of forgetting everything the moment a session ends. That matters. A lot of real work isn’t a single prompt. It’s a process that unfolds over hours or days, and until now agents were terrible at holding the thread.
The piece I care about more is Agent Identity. Google describes it as a native IAM identity type that enforces least-privilege permissions, logs every operation for non-repudiable auditing, and binds credentials to the agent’s runtime so a stolen token can’t be replayed from somewhere else. In plain English: the agent gets its own badge, that badge only opens the doors it needs, and every door it opens gets written down.
On the consumer side, Gemini Spark got hands. As of July 30, 2026, Spark can control the Chrome browser running on your actual desktop rather than a remote browser in the cloud. Once you approve it, it uses your logged-in accounts and saved passwords to run errands. Google’s examples are the tedious stuff: scheduling apartment viewings or researching and starting a flight booking. Chrome shows a notification when the agent is active, and the agent hands control back to you for payments and other sensitive steps.
Full disclosure: I haven’t personally run the Gemini Enterprise Agent Platform or Gemini Spark’s Chrome mode. I use the Gemini chatbot regularly, but the analysis below is based on Google’s documentation and reporting, plus my own experience deploying agents in an IT environment.
Why the identity part is the real news
Most coverage leads with the flashy demo of an AI booking your flights. I get it. But the quieter announcement, agent identity, is the one that actually determines whether any of this is safe to run at scale.
For years, agents borrowed a human’s login. That’s a governance nightmare. If the agent does something wrong, whose account did it? What did it touch? Can you prove it? Giving an agent its own least-privilege identity with a full audit trail is the difference between a tool you can defend to your auditors and one you have to explain away. I wrote about this same shift when Anthropic gave its Slack agent a dedicated identity in my piece on Claude Tag and why agent identity fixes a real problem. Google is now doing the same thing at cloud scale, and I think that’s the right direction.
An agent that can act needs its own identity, its own limits, and its own paper trail. The demo is the flight booking. The story is the badge.
What this means for you
If you run a small business, the Spark side is where you’ll feel this first. The promise of “hand off the tedious web errands” is genuinely appealing when you’re the whole staff. But read the fine print before you turn it loose: it uses your saved passwords and logged-in sessions. That’s your email, your bank tabs, your vendor portals. The safety design is reasonable, you approve payments and Chrome protects against prompt injection, but “reasonable” is not “zero risk.” Start it on low-stakes errands. Watch what it does. Don’t point a browser agent at anything you couldn’t afford to have it get wrong.
If you’re an enterprise IT leader, the Enterprise Agent Platform is your lane, and the questions are the ones you already ask about any service account. What can this identity reach? Is least-privilege actually configured, or did someone grant it broad access to save time? Are the logs going somewhere you monitor? Seven-day, stateful agents are powerful precisely because they persist, which also means a misconfigured one persists too. I’d treat every agent identity like a privileged service account, because that’s exactly what it is.
None of this is a reason to panic. It’s a reason to be deliberate. The same restraint that made AWS’s cloud-cost agent smart enough to refuse to delete anything is the mindset I’d bring here: let the agent investigate and prepare, keep a human on the trigger for anything irreversible.
My take
I’m genuinely optimistic about the enterprise identity work and cautiously curious about Spark. Google is building the boring governance plumbing that agents need before they belong anywhere near production, and I’d rather see a company lead with least-privilege and audit logs than with a splashy autonomy demo. The Chrome-driving Spark is the more exciting toy and the riskier one, mostly because saved-password access is a big blast radius in consumer hands.
The bottom line: agents are moving from advising to acting, and the winners will be the ones who get identity and permissions right first. Your next step this week is small and concrete. Go look at where your AI tools already have access, whether that’s a browser session or a cloud service account, and ask one question for each: if this went rogue for two days, what could it touch? If you don’t like the answer, tighten it before you add the next agent, not after.
News commentary by Brad Rowland — IT Infrastructure and Operations leader, automation builder, and AI implementer. Sources are linked inline.
