You don't learn Agent Studio by reading — you learn it by deciding what goes where, and being wrong safely. Below is the skeleton of a real production Fusion agent (identities removed, architecture intact). Your job: put the right kind of brain into each step. Then power it on and watch your agent run — including the moment it stops and waits for you.
THE SCENARIO
A warehouse user types into chat: “Ship 100 of ITEM-A and 50 of ITEM-B.”
Your agent must find every open purchase-order line for the supplier, allocate quantities oldest-first (FIFO), get a human's sign-off, and create a real Advance Shipment Notice in Fusion — without inventing a single fact.
What you just used — the 6 laws of shipping agents
🤖LLM at the edges only. Two AI nodes out of seven: understand language in, explain outcomes out. Everything between is deterministic.
⚙️Code for logic. Allocation is boring, auditable JavaScript — same input, same answer. An approver can trust it; an auditor can read it.
🔌APIs for facts. PO lines, sequence numbers, server time — fetched, never imagined. And REST paginates: production agents chain page-calls and merge, parsing shapes defensively.
👤Humans for accountability. Deltas not data-dumps, a change-request loopback, and a hard iteration cap. You felt this one yourself.
🪫Design the graceful no-op. If nothing can be allocated, the skip-flag bypasses creation and the report says so plainly. Agents must know how to do nothing.
📏The reporting prompt is a contract. Branch by status, totals not lines, shortfalls never hidden, hard length cap. An LLM with rules is a UI; without rules, a liability.
Your rank: —
Interview panels now ask exactly these design questions. Practice answering them under pressure — against interviewers who push back.
Based on a real production Agent Studio workflow with all customer-identifying details removed. Platform behavior evolves quarterly — see the Release Radar. · Companion reading: the AI Agents field guide.