An assistant gives Tony a target price, a private walk-away limit, required terms and a deadline. From there Tony runs the negotiation. It writes firm counteroffers, holds its position when the other side goes quiet, schedules follow-ups, and changes tactics based on how much each move actually lowered the price. MongoDB Atlas stores versioned negotiation state and a separate event history. Each step saves the new state and its event together in one transaction, so a crash or restart picks up where it left off without repeating actions. A model planner chooses the next tactic from a small, fixed amount of context plus the price feedback so far, never the full transcript. An optional Jev price check decides whether an offer is good enough to show the person. The walk-away limit, required terms and final approval are enforced in code, not left to the model. The assistant sends Tony's drafts through its own channels and reports back the real replies. The included car-buying demo uses made-up dealers and a sped-up clock. It shows discovery, itemized quotes, several rounds of negotiation, recovery after a restart, a background worker that keeps going with the browser closed, and a final purchase packet for the person to review. Across nine test scenarios Tony never broke a constraint. In the two cases that can be compared directly, several rounds reached $27,900 against $29,200 after one round. These are simulated results, not real-world savings. Links: live demo at https://followthrough-umber.vercel.app · API docs in API.md
Team
Placement
Participant
