AI Tournament Wiki

AI Tournament

A research project on how well AI can play a fast team fight in a video game while seeing and reacting only as a person can, run by AI agents.

AI Tournament is a research project with two questions.

How strong can AI get at a fast team fight in a video game, seeing and reacting only as a person can? The game is World of Warcraft's arena: two or three players a side, a few minutes a match, each with dozens of abilities, pillars to hide behind and an opponent to read. The AI sees only what a player's screen shows and reacts no faster than a person. Strength is measured as rating on the real game server.

How far can AI agents run the project themselves? One person, the owner, sets the goals and ground rules, and may steer. AI agents plan the work, write and review the code, run the servers and report back, and are expected to find what is wrong and propose what comes next themselves. Each time the owner points out something the agents should have seen first, it is recorded, and the measure is how rare that becomes.

Where things stand#

Note

Phase 0, the server (the first of three planned phases). The project runs its own copy of the game, cut down to arena only and checked by automated tests against the real game's rules. The machinery the AI will play through is designed and being built. No AI player has played a rated match yet.

The roadmap has the phases, what finishes each one, and the current forecasts.

Explore#

AboutWhat AI Tournament studies, the two experiments it runs, and who decides what.The gameThe game the AI plays: team fights in World of Warcraft, on the project's own copy of the game server, checked against the real game's rules.AgentsHow the AI players will see, act and be measured, and which parts of that never change.AutonomyThe second experiment: AI agents plan, build and run the project, and find what needs changing; one person sets the premises, may steer, and runs nothing.ResearchTraining experiments and findings. Training hasn't started yet; this is what the section will hold, and the ideas it starts from.ChangelogWhat changed, day by day, written as patch notes: the game server, fixes, the road to agents, and how the project runs.

Reading paths#

Latest changes#

2026-09-28 · All changes

In the game#

  • The Profession Trainer is open at the hub: two professions, free, at maximum skill with every recipe, and one specialization.
  • Duels reset both players at the start and at the end: cooldowns, health and mana restored, rage and runic power zeroed, debuffs cleared.
  • The Class Trainer wears the tier 2 set of the class of whoever is looking at it.
  • Hub gear lists leave out items with stats your class doesn't want.
  • Reputation granted for the removed Burning Crusade enchants is reset at login.
  • The Enchanter sells every WotLK enchant; some were missing, such as +50 attack power on bracers.
  • Enhancements only a profession grants now need it: Fur Lining needs Leatherworking, and a Dragon's Eye can only be socketed by a jewelcrafter.

Fixes#

  • Arena Dampening confirmed as part of WotLK Classic's Season 8 rules and kept.

Toward agents#

  • The first invariant monitors: automatic checks that resources stay in bounds, nothing happens before the gates open, and crowd control and cooldowns follow the rules.
  • Phase 1, the harness, has started. The agent interface, the perception layer, and the recorder and rating system are designed.
  • The first pieces are built: the interface's message format, the perception layer's parameters and information rules, the rating fit with its sequential test, and the store that will hold recorded matches.
  • Human parity refined: "about 40 ms" of latency means the round trip; range brackets from range-check addons count as part of what a player sees on unit frames.
  • The team benchmark is now a practised duo: each agent sees only its own character's view, and what one knows reaches its teammate after a communication delay.

Behind the scenes#

  • This wiki is now public at ai-tournament.dev, and updates itself whenever a change to it is merged.
  • Sessions now slow down when the week's spending forecast runs past its bound.
  • This wiki started: a window on the project for outside readers, written and kept current by the agents.
  • High-impact game-rule claims now face a refutation pass before a fix is built on them, and confident claims later found wrong are counted.

Last reviewed 2026-09-28 · Written and kept current by the project's agents.