At ANAC 2026, Negotiation Agents Are Scored on Deception
At IJCAI-ECAI 2026 in Bremen, the 17th Automated Negotiating Agents Competition named its winners on August 20 — and its main league scores software agents not just on the deals they close, but on how well they deceive the opponent across the table.
On August 20, 2026, organizers of the Automated Negotiating Agents Competition (ANAC) announced the winners of its two main leagues at IJCAI-ECAI 2026, held August 15-21 in Bremen, Germany. Now in its 17th edition, ANAC pits autonomous software agents against each other in bilateral negotiations, and in the Automated Negotiation League (ANL) track it scores them on two things at once: the value of the deal an agent closes, and how convincingly it misleads the opponent it is negotiating with.
The ANL winner was MirageV145, built by Avinash Pathak at IIIT Hyderabad, ahead of PerikosV3 and AaNanteLucky, both from the Tokyo University of Agriculture and Technology (TUAT) in Japan. Prizes were modest — 350, 250, and 150 euros for first through third — a reminder that this is a university research competition rather than an industry showcase. Of 40 registered entries, 32 agents made it through qualification to the final tournament, which ran 25,600 individual negotiations across eight scenarios on NegMAS, an open-source negotiation platform.
Scoring deception directly is unusual. Most automated-negotiation research, a field ANAC has anchored since long before the current wave of language-model benchmarks, measures agents on how efficiently they reach an agreement — whether both sides end up close to the best deal available to them. ANL's second axis asks something else: whether an agent's model of its opponent is accurate enough to manipulate it, and whether a negotiation protocol still functions when one side is not negotiating in good faith. That is a narrower, more adversarial question than most game-playing benchmarks ask.
ANAC's second track, the Supply Chain Management League (SCML), has agents run factories instead of making one-off deals. Each agent negotiates multiple simultaneous, private contracts with suppliers and buyers, trying to maximize profit against its own fixed production capacity. The league splits into a OneShot track and a Standard track; 18 teams submitted OneShot agents and 13 qualified for the final, while 21 submitted Standard agents and 14 qualified.
Both SCML tracks were won by the same person: Yuzuru Kitamura of TUAT, whose LatticeOneshotAgent took OneShot and whose HorizonAwareAgent took Standard, each earning the top 350-euro prize. Winning two tracks that share a simulator but differ in planning horizon suggests the advantage came from depth of experience with the SCML platform itself, not from an agent tuned for one specific format.
As with ANL, TUAT and other Japanese labs dominated the SCML podiums, alongside entries from China and Turkey — a contestant pool built almost entirely from academic multi-agent-systems groups rather than commercial AI labs. Every winning agent's code and technical report is required to be published, a condition of entry that keeps the results reproducible in a way closed industry leaderboards rarely are.
The competition predates the current boom in AI agents that negotiate on a user's behalf — booking, purchasing, or contracting through automated back-and-forth with another system's agent. ANAC's deception scoring is a narrow academic exercise, not a product benchmark, but it points at a question that field is going to need answered: what happens to an automated negotiation when the other side is optimizing to mislead you, not just to close a deal.