Implementing Agentic AI Ep 3: The Mission Brief

The HR team gave their new onboarding assistant a simple instruction: "handle the paperwork." Tuesday, nine in the morning: a hiring manager emails, "we agreed a higher figure with Sara — can you sort the letter?" The assistant, being helpful, drafted a revised offer letter with a salary nobody had approved — and queued it to send. It was caught in the outbox. By luck, not by design. "Handle the paperwork" turned out to contain a compensation authority nobody meant to give. Today: the three sentences that would have made luck unnecessary. The mission brief. Part three of five. For anyone about to hand real work to an agent — and for everyone who'll be accountable when it acts. In five minutes: the three-part brief that makes an agent safe to employ — goal, boundaries, escalation — how the same page decides which parts are workflow and which are agent, and the twenty-twenty-six toolbox that turns the brief into a running system. We're in HR, where the stakes are people, which is why the brief was invented. Every agent that touches real work needs three sentences written down before it starts. The goal — what done looks like, concretely. Not "handle the paperwork": "prepare contracts, equipment orders and first-day schedules for approved hires, matching the approved offer on file." The boundaries — what it must never touch, even to be helpful. Compensation figures. Employment terms. Anything legal. And the data it can even see is a boundary too: an onboarding assistant needs this hire's file, not the whole personnel drawer. The escalation — where it stops and asks. "Any request that changes an approved term goes to a human, always." Skip that third sentence and the agent does the only thing left when it's unsure: it guesses. Helpfully. In writing. Goal. Boundaries. Escalation. Three sentences on Tuesday morning, and the outbox story never happens. The same three sentences run the whole building. Finance, last episode: goal — match invoices to purchase orders; boundary — never approve a payment; escalation — mismatches above a threshold go to a person. Support: goal — resolve delivery queries; boundary — refunds above fifty dollars need a human; escalation — any legal-sounding complaint, straight up. Procurement, next episode: goal — track contract renewals; boundary — never sign anything; escalation — unusual terms go to legal. One page. Different nouns. That's what a method is. Now watch the brief do engineering. Read the onboarding goal again and episode one's question falls out of it: who chooses the next step? Contracts, equipment, schedules — the path is known. Workflow. Fixed stations, model inside them. But "answer new joiners' questions in their first month" — that's open. The path can't be drawn; an agent, on a leash the brief already wrote. One process, split correctly into both kinds — because the brief made the decision visible instead of accidental. So who actually builds this page into a system? Here's a secret worth the whole episode: every vendor diagram you'll ever see is three boxes wearing a costume — and that costume is how teams end up buying the same three boxes twice, at costume prices. Box one: a model that reads, reasons and writes — Claude is our example throughout this series. Box two: connectors, so it works from live systems instead of pasted screenshots — that's MCP, the open standard the industry now shares; your HR system, your document store, your calendar, plugged once. Box three: orchestration to run the fixed paths — LangGraph leads production use, with CrewAI and OpenAI's Agents SDK in the same family. Model, connectors, orchestration. Now you can read any vendor's diagram. One caution before you plug anything in: every connector widens what the agent can read — and an agent that reads documents can be given instructions by documents, planted by anyone who can write to them. Hold onto that; it's one more reason next episode's gates exist. A word about boundaries, because teams under-write them out of optimism. In HR the stakes are people: an over-helpful agent doesn't just make an error, it makes a promise — a salary, a start date, a visa assurance — that a human then has to unmake, to a person. So boundaries aren't distrust of the agent. They're respect for the people on the other side of it. You write "never touch compensation" not because the model is bad, but because some sentences should only ever come from a human. Hold that thought — it becomes the whole of episode four. Pocket rule number three: If the brief isn't written, the agent is writing it for you. Every unwritten boundary is a decision you've delegated by accident. The onboarding assistant didn't disobey — it obeyed a brief no one had written, assembled from helpfulness and guesswork. Three sentences. Written. Before it touches anything. Row three, then. Process: new-joiner onboarding. Goal, boundaries, escalation: written, on the page, in ink. Architecture: the split we drew — workflow for the fixed path, one leashed agent for open questions. Tools: model, three connectors, one orchestrated pipeline. Owner: the HR operations lead, by name. The playbook is halfway full — and notice it's still mostly sentences, not software. That ratio is correct. Three sentences, four takeaways. Every agent gets three sentences before it starts: goal, boundaries, escalation. The brief decides the architecture — fixed paths become workflows, open paths become leashed agents. The toolbox is three trays: model, connectors, orchestration. And boundaries aren't distrust — in HR especially, they're how you respect the humans on the other side. Next time: episode four — guardrails and the review ritual. Down to procurement, where an agent tracking contract renewals is about to do exactly what it was asked — and nearly renew the one contract everyone wanted dead. The approval gate that catches it is the whole episode. You'll learn the four-beat ritual — propose, approve, execute, review — and the four words auditors will ask about: owner, intent, scope, lifecycle. Five minutes. Bring a pen for that one.