At Google's Gemini at Work event on Thursday, the company announced its assistant is becoming an agent — one that plans work, uses tools, and connects to business systems on your behalf. Most coverage will focus on the feature list. One phrase from Google Cloud CEO Thomas Kurian deserves more attention than every bullet under it: the new agent can be given "objectives, not just instructions."
That's a five-word job description for the skill small business owners actually need next. Not prompt-writing. Delegation. And delegation has rules most people learn the hard way — an employee-shaped set of rules that now applies to software.
The difference between an instruction and an objective
An instruction is you doing the thinking and the agent doing the typing:
"Write a follow-up email to the contractor who asked about a quote on Tuesday."
An objective is you defining the outcome and the agent planning the work:
"Every quote that's gone unanswered for two weeks gets a yes or a no by Friday. Drafts queue for my approval. Anyone who's gone quiet twice gets flagged for a phone call."
The second is where the leverage lives. But here's the catch: most people hand an agent an objective it cannot possibly meet, because the objective is missing the three things an agent can't infer on its own. The agent plans — it doesn't read minds.
Three things every objective needs
1. A definition of done. "Improve our follow-up" is a wish. "Every stale quote gets a yes or a no by Friday" is an objective — there's a moment where you can check whether it happened. If you can't tell, from the outside, whether the work is finished, the agent can't either. Write the done-state first, in one sentence, with a number or a date in it.
2. Boundaries. What the agent may touch, and what it must never touch. The useful version is short: it may read the quote log, the email history, and the calendar. It may draft outbound messages. It must not send anything without approval, touch anything involving money, or contact anyone new for the first time. Two lists, five minutes to write, and the difference between an assistant and an accident.
3. An escalation rule. When to stop and ask. The one that works: "If you're uncertain whether something fits the boundaries, or if a customer replies with something angry, legal, or confusing — stop and bring it to me." Agents fail worst in the gaps between your rules. Give them an explicit place to stand in the gap.
If your objective has all three, the agent can plan against it. If it's missing any one, the agent will fill the gap with something — and you won't like what it fills it with.
Never let the first plan run unaudited
Google's announcement describes the mechanics plainly: you assign the agent work, schedule tasks, or let it respond to events, and it plans the work itself — creating a roster of sub-agents, each with its own identity, to run multi-step tasks in parallel. Treat that as the pattern, whatever tool you end up using: when the agent can plan, your job stops being writing steps and becomes reading plans.
The rule that makes objective-mode safe: the first run of any objective is a proposal, not a decree. Before the agent executes, read its plan. What will it touch, in what order? What will it send, to whom? Which sub-tasks did it invent you didn't anticipate?
Most first plans are wrong in a boring way — it'll prioritize the wrong customers, or draft in a tone nobody in your town uses. Catching that costs five minutes. Letting it run costs a week of cleanup and a customer who got a weird email. After the first plan survives review, later runs need only a spot-check: twenty minutes, once a week, reading what actually went out.
The agent signs its own work
One detail from the announcement matters more than it looks: Google's coworker agents get their own Workspace account — their own email address, calendar, Drive, and a presence in the company directory — and an audit trail attributed to the agent rather than to a person. Colleagues work with it the way they work with anyone else.
Strip the branding and that's an accountability architecture. When a customer asks "why did I receive this?", there's an inspectable answer: which objective, which rules, which approval. If you build a delegation habit on any stack, keep that property. Every agent action should trace back to a written objective and boundaries a human approved. That's what makes delegation defensible instead of deniable — and it's the part that survives the vendor changing everything else.
The model is a component, not a commitment
One more announcement detail worth sitting with: by default Google's agent picks the best model for the task, and users can override that — including choosing third-party models, starting with Anthropic's Claude. Google's agent, running Claude's models, connecting to Slack, Microsoft 365, Jira, and Postgres.
The practical read for a small business: the logo on the door is no longer the model doing the work, and it never will be again. Stop optimizing for picking the "right" model. Invest in the layer that outlives every model — the written objectives, the boundaries, the plan-review habit. Those transfer to whatever ships next quarter. There is always a next quarter.
When to use objective mode — and when not to
Objective mode earns its keep on recurring work with a clear done-state: follow-up cycles, review responses, report assembly, inbox triage, scheduling rounds, vendor paperwork chases. Instruction mode is still right for one-offs, novel situations, and anything where nobody has written down how the work is actually done. You can't delegate an objective for a process that only exists in someone's head.
That last point is the real adoption sequence: pick one recurring workflow, write down how it's done today, turn it into an objective with all three parts, review the first plan it produces, then let it run with a weekly spot-check. One workflow. That's the entire plan — and it's the same plan whether the agent arrives from Google, Microsoft, or a $20 automation tool.
The AI Agent Owner's Playbook covers this end to end — the objective-writing format, the permission ladder, and the plan-review checklist: The AI Agent Owner's Playbook — $49 CAD, instant download.
Top comments (0)