Google calls it the agentic era. What that phrase means for an office
Gemini 2.0 Flash arrived on 11 December 2024 with native image and audio output and the ability to call tools. The label is marketing; the shift underneath is not.
On 11 December 2024 Google introduced Gemini 2.0 Flash, which generates images and audio natively alongside text, can call external tools and services, and carries a very large context window. Google framed the release around what it called the agentic era.
That phrase has been used loosely all year. It is worth defining plainly, because the distinction determines what you can safely let a tool do.
Three levels, in plain terms
A chatbot answers. You ask, it replies, nothing happens in the world. Every risk is a wrong answer a person can catch.
A tool-using model fetches. It searches, reads a document, queries a system, then answers. Better grounded, and now the question of what it may read starts to matter.
An agent acts. It is given a goal and takes multiple steps on its own, possibly changing something: sending a message, updating a record, booking a slot. Here the question is no longer what it may read but what it may do, and whether you would notice.
Where the line should sit in an SME
Levels one and two are ordinary software decisions. Use them, scope the reading access sensibly, move on.
Level three deserves a rule, and the rule is simple: an agent may prepare anything and complete nothing that is irreversible. Drafting a reply, sorting incoming mail, assembling a quotation from your price list, flagging a discrepancy — all fine. Sending, paying, deleting, publishing, committing to a customer — a person confirms.
This is not caution for its own sake. It is the same principle as dual signature on a payment, which no Swiss company finds controversial.
The practical reading of this month
Between Microsoft’s agents at Ignite, Anthropic’s connection protocol and now Google’s framing, every major vendor is pushing in the same direction. You will be offered agents in 2025 whether or not you went looking.
The companies that handle this well will not be the ones with the best model. They will be the ones who decided in advance which actions require a human, and wrote it down before the first agent was switched on.