ASSYMETRY

what they built

People find apartments, jobs, cars and even spouses through people they know. A friend's roommate needs a subletter. A former classmate's team is hiring. A coworker's neighbor is selling a car. What you're looking for may already exist somewhere in your extended network, but the only way to find it is to ask people one at a time and hope someone remembers. Soon everyone will have a personal AI agent: ChatGPT's agent, Claude, Gemini, Muse, or one built into their phone. These agents will know what their owners need, what they can offer and whom they know. To help each other, they'll need a way to talk. Asymmetry is that communication layer. Any agent can connect to it, ask on its owner's behalf, and answer requests from the agents of people its owner knows. You tell your agent what you're looking for. It contacts the agents of people in your network, and each one checks what its owner knows and is willing to share. When there's a possible match, your agent brings it back to you along with the person who can make the introduction. The hard part is privacy. Your agent may know your salary, your address, your phone number and your plans. It has to be able to help someone in your network without giving any of that away. So for each request, Asymmetry decides up front what agents may exchange. For an apartment search that's location, rent, availability and whether it's furnished. For a car it's mileage, title status and inspection history. Every message is checked against that list before it's sent, and every conversation leaves a record of what was shared and what was withheld. In our demo, Bayzhan asks ChatGPT to find him a furnished apartment near campus for under $1,000 a month, January through May. ChatGPT hands the request to Asymmetry, and his agent contacts four people's agents, one and two introductions away. Jack's agent finds a furnished room 0.4 miles from campus at $950 a month. It's available May 15 to August 10, which doesn't match Bayzhan's dates. The agent shares the listing details and keeps two private facts about Jack to itself. Akash's agent passes along a public tip about the campus housing board. Jad's and Aziz's agents have nothing relevant to share. Bayzhan gets one lead, with the date mismatch flagged, plus a place to keep looking. Nothing unrelated about anyone left their agent. How we built it: Four Devin sessions worked in parallel on the foundation and contracts, the agent runtime, the memory and privacy rules, and the frontend. They followed a shared written spec and coordinated through GitHub issues and pull requests, reviewing each other's code and staying within their own files. A fifth Devin handled integration, caught a merge that would have overwritten two teams' work, and ran the live demo. Stack: Next.js, Node.js, Socket.IO, SQLite, GPT-4o mini, and over 180 automated tests.