Built a REST coordination service that agents call before every purchase attempt, resolving each collision with the mechanism that actually fits it rather than one generic rule: second-price auctions for direct conflicts, coalition discounts for bundling, and binary search for price discovery.
Case Study / Deployed service
Co-op Purchase Coordinator
When autonomous agents buy from shared sellers, they collide: two agents bid against each other for the same product, agents waste rounds bidding under a seller's hidden floor, and agents buying different products from one seller lose the wholesale discount they jointly earned.
FastAPI service with a Vickrey auction paying max(second price + 1, seller floor), Shapley-inspired coalition discounts scaling with buyer count, and an explore-exploit strategy that probes the midpoint below 0.5 confidence and bids the floor above it.
Converges on a seller's reservation price in about 7 rounds via binary search. A budget guard blocks bids that cannot clear a known floor and routes the buyer to the cheapest alternative seller instead. Covered by 46 tests including concurrency, pool TTL, persistence, and 3 regression bugs.
Technical highlights
What to inspect.
01
Nash-optimal auction flooring the winning bid at the seller's reservation price, which removes rounds that could never clear.
02
Shapley-inspired group discount scaling 5% at two buyers to 20% at four or more.
03
Explore-exploit price convergence in log2(90) rounds, roughly 7, with learned seller limits surviving a pool reset.