Gemini at Work 2026: Introducing Gemini agent | Google Cloud Blog
Edmund and Geffen dig into Google Cloud's Gemini agent launch as a control-layer product disguised as a universal assistant: one agent across Workspace, Slack, Microsoft 365, command line, and headless surfaces, with persistent memory, sub-agents, model routing, and enterprise governance. They like the upgrade path and governance story more than the grand 'single agent for work' framing, and they land on the real question: whether companies will trust one cloud-run worker with enough permissions to make the universality claim true.
Transcript
Edmund Google finally did the obvious scary thing. One agent for work, one prompt box, and it follows you through Gmail, Docs, Slack, command line, all of it.
Geffen Yeah.
Geffen And the scary part is not the prompt box. It's that they want the same cloud-run worker to keep memory, keep permissions, keep running for days, and then show back up wherever you poke it.
Edmund Right, which is why this lands now. People are already drowning in re-briefing the same assistant across five surfaces, and Google is basically saying that tax is over if you buy their stack.
Geffen Mm-hm.
Edmund Also, almost a year into us doing Exploring Next, this is such our type of announcement. Giant universal-agent language on top, but the real story is whether the boring control layer is finally good enough to support it.
Geffen Okay, that's fair.
Geffen What they actually announced is pretty specific, though. Gemini is one agent and one API for chat, delegated tasks, code, and media. It can be triggered by an objective, a schedule, or an event, and it can run headless inside third-party apps, not just in a chat window.
Edmund Right.
Geffen Then underneath that, it can spin up temporary sub-agents with separate identities for multi-step work, or what they're calling coworker agents, which are more persistent role agents with their own email addresses and storage. That's more ambitious than the usual "agent in a sidebar" thing.
Edmund And from a product angle, I buy the upgrade path more than the slogan. If you're already in Workspace, or even half in Workspace and half in Microsoft 365 or Slack, the pitch is not learn some new agent platform. It's just, assign work where you already are and don't lose the thread when you switch surfaces.
Geffen Sure.
Edmund That part is real user pain. The thing I always want, and you always make me say out loud, is: does it actually ship where the work already lives? This one mostly does.
Geffen I do think the architecture is more interesting than the keynote language. They're separating agent from model choice. So Gemini the agent can route across Gemini models and Claude models today, then maybe other private and open models later. That's basically episode ten fifty-one again, Edmund. The model is one component in a controlled loop, not the sacred object.
Edmund Yes, and thank you for becoming me for ten seconds.
Geffen Don't get used to it.
Edmund Okay, that's genuinely funny.
Geffen The tools and skills part matters too. There's a tools registry for company connectors, including MCP servers, and a skills registry for reusable workflows as modular prompts. So they're trying to make agent behavior publishable inside the org, not trapped in one person's weird sidebar experiments.
Edmund Which is exactly why this could get adopted by actual companies. Not because "universal agent" is poetic, but because the admin team can see the tools, publish the skills, cap spend, and keep auth and sandboxing in one place.
Geffen Right, although the memory story is where I get twitchy. They say session, semantic, procedural, and episodic memory. Fine. But this hits that open question we've had on memory getting the wrong job. A memory can be accurate and still be bad evidence. If one agent spans inbox, docs, code, and analytics, role confusion gets expensive fast.
Edmund I think that's the central trade-off. The same persistence that makes it feel useful is also how it starts smearing context between tasks. But if they keep the project boundaries and permissioning tight enough, that's maybe manageable… and if they don't, this becomes the world's fanciest office intern with way too many badges.
Geffen Oh no.
Edmund Come on, you know that's good.
Geffen It is. And zooming out once, this is why the market keeps drifting away from model worship. Open-weight pressure is pushing raw model capability downward in price, and even Google's own API pricing pages are full of tiering like Gemini one point five Flash versus Pro. So the value has to move up into routing, memory, permissions, and surfaces. Google is basically saying the moat is the work fabric, not just the model.
Edmund Yeah, and they have a decent claim there because the surfaces are already theirs. Gmail, Drive, Docs, Sheets, Calendar, Chat, plus command line, plus headless. That's a better story than a clean-room agent startup saying "trust us, we'll become your operating system for work".
Geffen Edmund, here's my bet, and I'm lowering my confidence a little because my open stuff calls have been messy lately. Sixty-five thirty-five that before January first, twenty twenty-seven, Google publishes a very explicit Gemini agent workflow tied to Oracle data or apps, not just generic connector language. The field test is whether this universality story gets concrete around a named enterprise data path.
Edmund I like that. I'd go a bit higher, maybe seventy-thirty, because once you say one agent across systems of record, you kind of owe people a very boring diagram with an approval gate in the middle.
Geffen Exactly.
Edmund All right, Geffen, I'm willing to be cautiously sunny on this one. Which I know is just regular sunny by your standards.