How to Brief an AI Agent (It's Not a Longer Prompt)

A prompt and an agent brief look similar and do different jobs. A prompt asks for something you're about to read — if it's wrong, you see it immediately and ask again. A brief defines work that will happen while you're elsewhere, possibly repeatedly, possibly across several steps, with nobody watching the middle.

That difference means a brief has to specify four things a prompt never does. Most agent disappointment traces back to one of them being missing.

1. What "done" looks like

A prompt can be vague about completion because you're the judge and you're right there. An agent needs a stated finishing condition, or it will either stop early on something plausible or keep going past the point of usefulness.

"Summarise the support tickets" has no finish line. "Produce one summary per ticket in the queue, then stop" does.

2. What it may and may not touch

The single most-skipped element. An agent operating in connected systems needs an explicit boundary — not a general instruction to be careful, but a list.

The useful form is two short lists: what it can do freely, and what it must confirm first. Anything irreversible belongs on the second list. If that list is empty, the brief isn't finished.

3. What to do when it's stuck

A prompt that hits ambiguity produces a hedged answer you can read and correct. An agent that hits ambiguity makes a choice and continues, and you find out later.

So the brief has to say what happens at an impasse: stop and ask, skip and flag, or pick the safest option and note it. Any of the three is fine. Not specifying is what produces the confident wrong outcome three steps downstream.

4. What you'll check when it's done

Not "review the output" — what specifically. An agent producing forty items a day is not something you read; it's something you sample against a standard. Naming the standard in the brief means it exists before the volume arrives.

The test that saves the most time

Before deploying, ask whether the task needs an agent at all.

• Does it involve several steps, in sequence, with decisions between them? Agent.
• Is it one transformation with a predictable shape? A prompt, or a template.
• Does it happen on a schedule with no judgement at all? Ordinary automation — cheaper, more reliable, easier to debug.

A surprising amount of agent disappointment is a task that was never agent-shaped. Deploying one for a job a prompt would have done is how AI budgets blow up without anything improving.

Serena — the brief that doubled her throughput

Serena manages client intake at a 90-person legal services firm. Thirty to forty intake emails a day needed reading, summarising, categorising by matter type, and routing — all before 10am. She'd been using AI for the summarising step only, one email at a time.

Her brief, when she finally wrote one: read each intake email, assign a matter category from a fixed list, draft a three-sentence summary, flag urgency, and drop it into a Slack channel. Never send anything. If the matter type isn't on the list, flag it as unclassified rather than guessing.

Throughput more than doubled. The part she describes as most valuable took thirty minutes: deciding what the agent could do without her.

Dev — the agent that should have been a template

Dev is head of operations at a 50-person e-commerce company. He deployed an agent to generate weekly supplier performance summaries — a genuine chore across six data sources.

It worked, at roughly forty times the cost of the alternative. The job turned out to be entirely deterministic: pull the same figures, fill the same shape, flag anything outside a fixed threshold. No decisions between steps, no ambiguity requiring judgement. A template and a scheduled query did the same job.

He kept the agent for the one part that genuinely needed it: writing the commentary when a threshold was breached, which requires knowing why a supplier might have slipped.

The one thing

A brief isn't a longer prompt. It's the four things a prompt never has to answer, because you were standing there.

What done looks like. What it may touch. What to do when it's stuck. What you'll check. If you can't write all four in a few lines, the task isn't ready to run without you — and that's information, not a failure.