Google shifts Gemini into an agent that takes objectives, not just instructions

Gemini graduates from assistant to agent

At its Gemini at Work event on Thursday, Google said Gemini is moving beyond an assistant that waits for prompts: it is becoming an agent that plans and carries out work across a company's own tools and systems. The details come from a single account published on dev.to, so the specifics should be read as reported rather than independently confirmed.

According to that post, the framing rested on a phrase from Google Cloud CEO Thomas Kurian: the new agent accepts "objectives, not just instructions." In practice, users can assign the agent work, put tasks on a schedule, or let it react to events, and the agent does the planning itself — including creating sub-agents, each with a separate identity, that run multi-step tasks in parallel.

What objective-based delegation means

The dev.to post draws a sharp line between the two modes. An instruction has the human doing the thinking and the agent doing the typing — write this email to that person. An objective defines an outcome — every unanswered quote gets a yes or a no by Friday — and the agent works out the steps to get there.

The author's argument is that this turns prompt-writing into delegation, and delegation needs structure. Objectives fail, the post says, when they lack three things the agent cannot infer: a definition of done that can be checked from the outside, boundaries listing what the agent may touch and what it must never touch, and an escalation rule telling it when to stop and ask a human. Missing any one of them, and the agent fills the gap with its own guess.

The post also prescribes a review discipline: the first plan an agent produces should be treated as a proposal that a human reads before anything executes, with subsequent runs reduced to a short weekly check of what actually went out.

Agents get their own accounts

One detail the post flags as underappreciated: Google's coworker agents receive their own Workspace accounts — a dedicated email address, calendar, Drive storage and a listing in the company directory — with audit logs that attribute actions to the agent itself rather than to a person.

Strip away the branding and this amounts to accountability built into the design, the author argues. When a customer asks why they received something, there is an inspectable answer: which objective drove it, under which rules, with which approval. That property is what makes delegating to software defensible, and the post advises keeping it no matter which tool is used.

The model becomes a swappable part

The announcement also loosens the tie between Google's brand and Google's models. By default, the agent chooses whichever model suits a given task, and users can override that — including picking third-party models, starting with Anthropic's Claude — while the agent connects to systems such as Slack, Microsoft 365, Jira, and Postgres.

The dev.to author's takeaway for businesses: a vendor's logo no longer tells you which model is doing the work, and it will not again. Choosing the "right" model matters less than investing in the layer that outlives every model — written objectives, boundaries, and the habit of reviewing plans.

The post adds a note on fit: objective mode suits recurring work with a clear finished state, such as follow-up cycles, report assembly or inbox triage, while instruction mode remains appropriate for one-offs and processes that only exist in someone's head.

Why it matters

Two shifts are packed into one announcement. First, the human role moves from writing steps to reviewing plans, which turns delegation hygiene — done-states, boundaries, escalation rules — into a mainstream workplace skill rather than a prompt-engineering trick. Second, giving agents their own accounts and audit trails sets a template for accountability when software acts on a company's behalf. And with model selection becoming a setting rather than a commitment, the durable investment is the process layer around the agent, not any single model behind it. One caveat: this rests on a single commentary source, and fuller coverage should fill in availability, pricing and limits.