Commerce operations
Product research, listings, customer support drafts, and order follow-up for an online store.
Scenario
An online store — dropshipping or your own inventory — is a stream of repetitive, interruptible work: find products worth selling, write listings that convert, answer the same support questions, chase orders and reviews. Every task is individually small; together they consume the operator. This playbook staffs the stream.
The team
| Agent | Job |
|---|---|
| Product scout | Researches niches, trends, and suppliers with the web tools; maintains a scored candidate list |
| Listing writer | Turns a chosen product into title, copy, and variants — in the store's voice, SEO-aware |
| Support agent | Drafts replies to customer messages from the policy docs and order context; a human approves and sends |
| Ops agent | Order status checks, review follow-ups, weekly margin and stock readouts |
The wiring
- The store's knowledge is a synced source: policies, supplier notes, product sheets, past listings. The support agent quotes the actual refund policy because it can search it — not a hallucinated one.
- Store platforms attach where their interfaces are: an MCP server or public API becomes an agent capability via a skill with a scoped credential; for everything that is only a dashboard — supplier portals, marketplace seller centers — the machine desktop's persistent, logged-in browser does the clicking, with screenshots flowing back into chat for the audit trail.
- The candidate list, the listing queue, and the support drafts are
plain files in
data://— reviewable in the Files widget, searchable by every agent, versioned on every write.
The rhythm
- Daily 07:00 — Product scout: refresh the candidate list, flag movers.
- On demand — Listing writer: you pick a candidate, it delivers the listing package for review.
- Every 2 hours — Support agent: sweep new messages into drafted replies.
- Daily 19:00 — Ops agent: order exceptions and follow-ups; Sunday — the weekly margin readout.
Governance
Money and customers are involved, so the defaults are conservative: drafts, not sends — outbound messages and listing publishes wait for a human tap; supplier and platform credentials are project-scoped and never reach agents that don't need them; spend caps per agent keep research costs flat; and the audit log records every shop-touching action with the identity it ran under.
Extend it
Scale sideways into more stores (one project each — the isolation boundary keeps suppliers, credentials, and voice apart) or deepen the team: a pricing agent watching competitors, an ads agent feeding the marketing team playbook. The store grows by adding seats, not hours.