Objectives, Not Instructions: How to Write Work That AI Agents Can Actually Take Over

At Google's Gemini at Work event, the company announced that its assistant is evolving into an agent — one that can plan work, call tools, and connect to various business systems on your behalf. Most coverage will focus on the feature list, but one line from Google Cloud CEO Thomas Kurian deserves more attention than any individual feature: the new agent can be given "objectives, not just instructions."

Those few words neatly summarize the capability small business owners actually need to develop next. Not prompt-writing — delegation. And delegation has rules, ones most people learn the hard way. Rules that originally applied to employees now apply to software.

The Real Difference Between Instructions and Objectives

An instruction means you do the thinking and the agent does the typing:

"Write a follow-up email to the contractor who asked for a quote on Tuesday."

An objective means you define the outcome and the agent plans the work:

"Every quote that hasn't received a response in two weeks must get a clear 'yes' or 'no' by Friday. Queue drafts for my approval. Flag any client who's gone quiet twice — those need a phone call."

The second form is where the leverage lives. But there's a trap: most objectives people hand to agents can't actually be executed, because they're missing three elements the agent cannot infer on its own. The agent will plan — but it won't read your mind.

The Three Elements Every Objective Must Contain

1. Definition of Done. "Improve our follow-up" is a wish. "Every overdue quote gets a response by Friday" is an objective — there's a specific moment when you can verify whether it actually happened. If you can't tell from the outside whether the work is complete, neither can the agent. Write the completion state in one sentence, with a number or a date attached.

2. Boundaries. What the agent can touch, and what it must never touch. The practical version is short: it can read the quote log, email history, and calendar; it can draft outbound messages; but it cannot send anything without approval, cannot touch anything involving money, and cannot initiate contact with anyone it hasn't heard from before. Two lists, five minutes to write — that's the line between "assistant" and "incident."

3. Escalation Rule. When to stop and ask you. An effective version looks like this: "If you're unsure whether something falls within the boundaries, or if a client's reply contains anything angry, legal, or confusing — stop and hand it to me." Agents are most likely to go wrong in the gaps between your rules. Give the agent a clear landing spot for those gaps.

If your objective contains all three elements, the agent can plan against it. Leave out any one of them, and the agent will fill the gap with something — and you won't like what it fills in.

Never Let the First Plan Run Without Review

Google's announcement makes the operating model clear: you assign work to the agent, schedule tasks, or let it respond to events — then it plans the work itself, creating a set of sub-agents, each with its own identity, executing multi-step tasks in parallel. Whatever tool you end up using, treat this as the paradigm: when an agent can plan autonomously, your job is no longer to write steps. It's to read the plan.

The rule that keeps "objective mode" safe is this: the first run of any objective is a proposal, not an execution. Before the agent acts, read its plan. What will it touch, and in what order? What will it send, and to whom? What sub-tasks did it invent that you didn't anticipate?

Most first-draft plan errors are mundane — it'll prioritize the wrong clients, or draft emails in a tone nobody in your town would recognize. Catching these takes five minutes. Letting it run unchecked costs a week of cleanup and one client who got a strange email. Once the first plan passes review, subsequent runs only need spot-checks: twenty minutes once a week, reading what actually went out.

The sub-agent architecture is worth a closer look. In a multi-agent system, an orchestrator agent understands the objective, breaks it into tasks, and distributes sub-tasks to specialized execution agents — one to read the calendar, one to draft emails, one to query CRM records. This parallel structure dramatically speeds up execution, but it also means a single objective can trigger far more actions than you'd expect. You think you're sending one follow-up email; behind the scenes, three sub-agents may simultaneously be reading your inbox, querying quote records, and updating customer status. This is precisely why the first plan must be reviewed: what you need to see isn't a single action, but an action tree.

Agents Sign Their Own Work

One detail in the announcement carries more weight than it appears: Google's "coworker" agents have their own Workspace accounts — a dedicated email address, calendar, Drive, a presence in the company directory, and an audit trail attributed to the agent itself rather than to any individual person. Colleagues collaborate with it the same way they would with anyone else.

Strip away the branding, and this is an accountability architecture. When a client asks "why did I receive this email?", there's a traceable answer: which objective triggered it, which rules governed it, which approval authorized it. If you're building delegation habits on any tech stack, preserve this property. Every agent action should be traceable back to a written objective and a human-approved set of boundaries. That's what makes delegation defensible rather than deniable — and it's the one thing that survives after vendors change everything else.

The concept of accountability architecture originates in corporate governance and compliance, where the core requirement is that every decision or action can be traced back to a specific authorizing party, scope of authority, and triggering rationale. In traditional IT systems, this is typically implemented through access logs and permission management. Giving agents their own accounts and audit trails is essentially extending that mechanism from people to software actors. For small businesses, this isn't just a technical detail — if an agent sends a client a poorly worded email on your behalf, or incorrectly flags a receivable, "I didn't know it would do that" is not a viable defense. Being able to answer "which written objective triggered this action, and who approved it" is the only answer that actually holds up when a client complains or a dispute arises.

The Model Is a Component, Not a Commitment

Another announcement detail worth sitting with: Google's agent will by default select the most appropriate model for each task, and users can override that choice — including opting for third-party models, with Anthropic's Claude among the first supported. A Google agent, running a Claude model, connected to Slack, Microsoft 365, Jira, and Postgres.

The practical implication for small businesses: the logo on the door no longer equals the model doing the work, and it never will again. Stop obsessing over picking the "right" model. Put your energy into the layer that outlasts any model — written objectives, boundaries, and the habit of reviewing plans. These transfer to whatever ships next quarter. And there's always a next quarter.

When to Use Objective Mode, and When Not To

Objective mode delivers the most value on repetitive work with a clear completion state: follow-up loops, review responses, report rollups, inbox triage, scheduling rotations, vendor document chasing. Instruction mode still makes sense for one-off tasks, novel situations, and anything where how the work gets done hasn't been written down yet. You can't delegate an objective for a process that only exists in someone's head.

That last point is the real sequencing guidance: pick one recurring workflow, write down how it works today, turn it into an objective with all three elements, review the first plan it produces, then let it run and spot-check weekly. One workflow. That's the whole plan — and it works regardless of whether the agent comes from Google, Microsoft, or a $20 automation tool.