At its Gemini at Work 2026 event on October 8, Google Cloud announced the "Gemini agent," which it describes as a new single, universal agent for work. It is available inside the Gemini Enterprise app, Google Workspace, and third-party services, and Google frames the pitch in one line: you give it objectives, not instructions. It can answer questions, handle knowledge work, create media, and write and run code. It is currently in private preview, with wide availability promised soon for Workspace customers on select Business and Enterprise plans.
That framing matters more than the feature list. For two years, "AI agent" has meant a thing you assemble: a model, a prompt, a tool connector, a workflow you design. Google is now selling the opposite premise, one agent that plans the work, picks the tools, and picks the models, so that buyers stop building and start delegating. It is the strongest version yet of the universal-agent thesis, and it deserves a clear-eyed read.
What the Gemini agent actually does
The examples Google gave at launch, reported by 9to5Google and VentureBeat on October 8, are deliberately mundane, which is what makes them credible. If a manager emails asking for the latest project update as a slide deck, Workspace Intelligence recognizes that as a delegatable task and offers a single-click option to pass it to Gemini. You can ask the agent to set up a meeting with the usual team of regional event leads next week without supplying a single name or email address; it determines who those people are from the membership of your chat space and your last event's thread, checks their calendars, and starts an email thread to coordinate a time that works, including with external participants.
Beyond the personal-assistant mode, the agent can act as a "team member," working on behalf of a group, like a project manager within a team, or on behalf of a specific role, like an analyst in a finance department. It is reachable from any device and channel, with command line, Google Workspace, Slack, and Microsoft 365 given as examples, plus an API to embed it in third-party applications as a headless agent with no dedicated interface. In Gmail, Docs, Sheets, Slides, and Chat, you can invoke it with @Gemini.
Under the hood, the agent runs each job on the model that fits best. That currently includes both the Gemini and Claude families of models, with support for other private and open models planned, which Google says delivers optimal quality and lower cost. Sub-agents tackle multi-step tasks, connecting to third-party tools in workflows that can run in parallel and sequential steps for hours or days.
The four memories
The most technically interesting decision is persistence. The agent runs in the cloud over a single set of memories, context, and one personalization graph, with four distinct memory types Google describes explicitly:
| Memory type | What it holds | Why it matters |
|---|---|---|
| Session | The task in front of it, even over days | Long jobs survive restarts and interruptions |
| Semantic | A structured knowledge base built from documents and conversations | The agent accumulates your domain facts |
| Procedural | How a job gets done, including skills it writes for itself | It learns your processes, not just your facts |
| Episodic | Everything it has done before | Its track record is inspectable, in principle |
An agent that writes its own skills and keeps an episodic record of its own actions is a fundamentally different artifact than a chatbot with a long context window. It is also a governance object: procedural memory means the way work gets done drifts toward what the agent learned, not necessarily what your policy says.
Universal agent or tailored build: the honest comparison
Google's implicit argument is that model selection, orchestration, and memory are too complicated to leave to individual teams. There is a strong version of this argument and a weak one. The strong version: most businesses should not be stitching together model-routing logic, and a platform that does it well will beat a hand-rolled system on cost per quality. The weak version, the one to resist: that one agent can understand your business as well as a system built around it.
| Dimension | Universal agent (Gemini agent) | Tailored agent build |
|---|---|---|
| Time to first value | Hours, if you already use Workspace | Weeks, done well |
| Fit to your actual process | Learns what it observes | Encodes what you specify |
| Model choice | Google picks, per job, from Gemini + Claude | You pick, per job, from anything |
| Data gravity | Stays in Google's graph and stack | Stays in your architecture |
| Governance surface | Vendor-controlled, one personalization graph | Yours to design and audit |
| Cost shape | Subscription + usage, less controllable | Direct, fully attributable |
The honest middle position for most companies is hybrid: universal agents for the fluid knowledge work that fills calendars, purpose-built agents for the few processes that actually differentiate the business. That is the same pattern we described in Cohere North 2's enterprise agent control room coverage: platforms consolidate, but the governance layer decides what consolidation is safe.
What SMBs should actually take from this
If you run a small or mid-size business, the announcement is less about buying this specific product and more about what it makes normal. Three practical reads:
- The @Gemini interface will train a generation to delegate, not instruct. When your staff starts forwarding objectives to agents by reflex, the businesses that benefit are those with clean, documented processes. Chaos documents itself badly; good systems compound.
- Multi-model routing is now table stakes. Google routing to Claude models on October 8 is the quiet headline: buyers no longer need to marry a model, they need to own their requirements. Skill with models is becoming a procurement question, not an engineering one.
- Persistent, self-writing agents need an inventory. An agent with its own email address, calendar, and procedural skills is a systems actor. Ask three questions of any agent product you adopt: who can see what it remembers, who approves the skills it writes, and where its actions are logged.
Open questions we'd want answered before production use
Private preview announcements raise more questions than they answer. Ours: How exportable is the personalization graph, or is the memory a retention moat? What does the audit trail of agent-initiated emails and calendar actions look like to an external participant? And what happens to episodic memory when an employee leaves, or when the agent was acting on behalf of a role that changes hands? The four-memory architecture is elegant; the hard part of enterprise AI was never elegance.
There's also the competitive ledger. Microsoft, Anthropic, and OpenAI are all pushing toward the same place from different directions, and universal agents from any vendor concentrate enormous leverage in whoever owns the memory and the routing. If part of your differentiation lives in how you work, think carefully before all of it lives in someone else's graph.
Deciding what belongs in a universal agent and what belongs in a purpose-built system is the central architecture question of this cycle, and we help teams answer it.
Read next:
