StudioHackathon build: a buyer agent negotiating against 8 seller agents, and what went wrong
Original title: Built a buyer agent that negotiates against 8 seller agents. A few things that went wrong
For a hackathon the author built a group-buying system where one negotiator agent bargains against 8 seller agents, each with a private floor price and stock, submitting sealed offers. Money is handled by a deterministic engine rather than the models, and 200 random auctions ran with zero rule violations. Using Nemotron 3 Ultra (thinking off), failures included repeatedly calling already-answered read tools, messaging all sellers at once, ending turns with text instead of tool calls, and a smaller model returning prices as text like "$39.95"; fixes were removing answered tools, a fixed reminder message, and closing the string option in the schema. Every call uses a strict JSON schema with code-level re-validation, plus per-negotiation and daily spending caps that trigger a labelled replay…
The author shares four concrete failure modes in a multi-agent negotiation system, like looping on read tools, along with the fixes that worked.
The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.