The Fig AI Design Makeathon gave us five days and one brief. Design the future of human and agent collaboration. Show what a person needs to understand or decide, how they act, and how the system responds.
We started from the agent workflow scenario. A project's knowledge lives across meetings, documents, chats and tools. Plans change. Agents risk acting on information that's out of date, or that contradicts itself.
We took that into travel. A flight moves, a storm cancels a tour, a hotel loses a room. Agents catch the clash, propose the fix on the itinerary, and the advisor approves it.
Tandem is a workspace for travel advisors. Five of us built it: Anne Lin, Anna Hu, Sebastian Moreano Mesa, Victor Fang and me.
Accepting every change isn't control
In September 2025, 6,019 of 6,500 American Airlines flights moved. When a flight moves, the car and the plan after it have to move too.
Agents can already find a cheaper fare or rebook a hotel on their own. That only helps if the advisor sees everything a change touches before it goes out. Otherwise the advisor finds out what broke when the client does.
We didn't interview advisors. We watched them. 17 public videos, about 11 hours, of advisors sharing their screen while they work with AI. One advisor had spent two days clicking accept on every change their AI tool proposed.
"I fell asleep last night hitting accept. It has taken me two days."
6 in 10 US advisors have tried generative AI. The agents are already at the desk. The question is how to work with them.
Agents find, hold and draft. The advisor decides.
That became the rule for every agent in Tandem. A research agent finds options. A booking agent holds and books. A contact agent writes to the client. Anything that changes a booking, costs money or reaches the client waits for the advisor.
Take Saturday. Ana and Leo fly from Rome to Palermo at 08:40, and Marco, their guide, meets them at 12:00. On Thursday morning an agent finds a flight that's €212 cheaper. It lands at 13:05.
The agent doesn't decide. Only Ana can move Marco. So Tandem drafts the message to Ana, and nothing leaves until the ten-second count runs out or the advisor sends it. While they wait, the booking agent holds the seat.
Ana says yes. Marco offers Sunday. The flight, the car and the guide move together, and the trip drops from €5,632 to €5,420. One decision from the advisor, one answer from the client. If Ana says no, Marco says no or the hold runs out, the hold drops at no cost and the trip stays as booked.
Every price and time in the prototype is a sample number.
One change, shown across the trip
A change is never just one booking. So the board shows the whole trip by the hour, with dates across, bookings in rows and activities on an hour grid. A clash shows up as two things at the same height.
When a flight lands late, you see the late tag, the guide in red, the car that follows, and the price of the choice beside the total. Tools open on the thing they change, not in a panel somewhere else. The agents live in one side panel, Trip agents, with Work for what they're doing and Chat for asking them.
And every action can come back. Ten seconds of Undo, then Withdraw, with two lists: what comes back, and what stays. We turned down an Undo that pretends a sent message can be unsent.
A design system that fails the build
Tandem runs on one token file. 32 colours, 31 spaces, 11 type sizes and 7 corners, in Light and Dark. 268 components, each with its states, and 331 tests. A raw colour or pixel value fails the build.
A teammate drops one HTML file into an inbox folder and opens a pull request. The intake maps every value to a token, decides if it's a new component or a variant of one that exists, and reports back on the pull request.
We split the work with AI too. Gemini read the 17 advisor videos and pulled the quotes, and a script matched 208 of 234 word for word against the captions. Codex built the design system in code. Claude built the itinerary prototype and its flows.
3rd of 76 submissions
Tandem placed 3rd at the Fig AI Design Makeathon 2026, out of 76 submissions from 192 teams and 329 participants.
Five of us, nine meetings and 9 hours 15 minutes of calls, from Nov 7 to Nov 9. Each of us brought our own explorations, and we kept the parts that worked.
Some things are still open. Where the history of decisions should live. Whether ten seconds is the right length for Undo. A single "confirm all changes" step, raised on a call and not built yet.