Reeve

Our own product: an AI workforce with a person in charge

Reeve is a product we designed without a client: an operating system for an AI workforce at an operations business. Four agents cover operations, sales, support and finance. They notice, propose, act and report, and every action they take is approved, logged and easy to undo. The mock company is Northstar Retail, a business with shipments, carriers, invoices and a support inbox, which is where agent work gets real quickly. We sell it as a Figma kit for teams building AI agent and operations products.

The dashboard is not the hard part of an agent product. The handoff is: what an agent does on its own, when it stops and asks, what proof it shows, and how a person corrects it without starting over. Every screen in Reeve was designed around that handoff.

Agents with a job and a person they report to

Setup takes about fifteen minutes over six steps, and the third one introduces the team. Atlas runs operations and reports to Maya Patel. Nova handles sales for Marcus Reed, Mia answers support for Daniel Brooks, and Ledger checks the books for Sarah Chen. Supervisors approve their agent’s work and get its daily summary, so every agent has someone accountable for it from the first day.

Agents never wear human faces. Each one has a glyph and its own colour, and people always appear as real photos, so nobody has to wonder whether a person or an agent did something.

The morning starts with what needs a person

Home is a briefing, not a wall of charts. Atlas opens the day with one line: one late VIP shipment and two money decisions need you. Below it sit the numbers that matter, 47 tasks closed overnight and 3 that need a person, then the approvals waiting, each with its evidence and an approve button, and a list of what is still undoable.

One decision, end to end

At 09:24 Atlas sees a carrier push a VIP order back by 36 hours. It checks the carrier’s history, finds four of the last six shipments for that customer arrived late, and drafts an email with the new delivery date. The customer is a VIP, and company policy says a person reviews every external message, so the draft stops and waits for one of two named approvers.

The task page shows all of it: six stages from detected to verified with a time on each, every step the agent took with the records it used, the draft itself, and what happens after approval. The email goes out with a fifteen minute undo window, Atlas watches the new arrival time, and the task closes when the customer confirms. Detected at 09:24, verified at 10:58.

Autonomy per action, earned with a record

Autonomy has four levels, Watch, Suggest, Ask first, and Act and report, and it is set for each action rather than for each agent. Atlas can read carrier data and pause a few shipments a day on its own, while emailing a customer always asks first. Each allowed action carries its own limit and undo window. Some things no agent can ever earn: moving money above the policy limit, deleting a record, messaging people outside the company, or changing another agent’s settings.

More freedom comes from a track record, not a toggle. When 48 of Atlas’s last 50 reroute proposals were approved unchanged and none were undone, Reeve suggests letting it reroute on its own, for 30 days, up to 12 a day, with VIP customers kept on Ask first and a 24 hour undo window. The supervisor is told, and the grant can be revoked at any time.

Every task, and every way it can go wrong

The tasks board shows 128 open tasks in the stages they move through, and a column for the ones that need attention: a sync that failed with a retry button, a reconciliation running slow, a request a person undid with a link to what changed. The unhappy paths got the same care as the happy one, with screens for a task that partly failed, an action blocked by policy, a low confidence result and an approvals queue that is all clear.

Ask Reeve, with sources

Anyone on the team can ask a plain question, such as why one customer’s shipments are running late this month. Reeve answers in a sentence, then shows its working: on time rates by carrier, the numbered sources it read, and the next steps it suggests, like creating a workflow or opening the incident. An answer you can check is an answer you can act on.

A brand and a system built for agents

The file starts with the brand, then the foundations: 49 variables and 12 component sets with 67 variants, plus 24 standalone components, documented with the rules that keep agents and people visually distinct. The 76 screens run across twelve flows at 1440 wide, from onboarding and home through agents, tasks, approvals, incidents, operations, intelligence, workflows, control and settings, to a flow of edge states, with 98% of frames in auto layout. The kit is on our products page.