The question#
How well can AI agents play World of Warcraft: Wrath of the Lich King arena without ever perceiving or acting in ways a human player couldn't? (New to the game? See arena in five minutes.)
The far goal is optimal play for every team composition (comp): the best builds and play each comp allows, found within human limits. It is not a search for the single strongest comp; every comp keeps evolving and none is favoured. For teams, the benchmark is a practised duo: coordination as good as the comp's synergy allows, within human limits.
AI agents run the project themselves: they decide what to work on, short and long term, and do it, inside walls that keep the owner's machine safe. See autonomy.
The premises#
The premises are the part of the project that does not change. They are the owner's alone; agents may propose a change but never make one. Everything built on top of them is replaceable.
| Premise | What it means |
|---|---|
| Arena only | Rated 2v2 and 3v3 under WotLK Classic's final rules (Season 8), emulated on the 3.3.5a client and server engine, on Nagrand, Blade's Edge, Ruins of Lordaeron and Dalaran Sewers. See the game. |
| Gear baseline | Season 8 Wrathful gear including the item level 277 weapons, plus PvE items up to item level 264. See the 264 meta. |
| Human parity | Agents never perceive or act in ways a human player couldn't. See human parity. |
| Humans can play | A human can queue, and agents fill the remaining slots. |
| A strength gradient | Agents span from a weak baseline to the current best. |
| Budget | All game compute, training included, runs on one desktop PC: a Ryzen 7 5800X, 32 GB of RAM and an RTX 3060 Ti. Agent reasoning comes from a separate model subscription. |
| Autonomous operation | The owner owns the premises, the parity principles, the stable architecture, the measurement rules and the budget, and may steer. Agents own everything else, within limits the environment enforces. |
The stable architecture#
The parts that define the problem are fixed, so results stay comparable: the real game server, the perception layer, the agent interface, the format for loadouts and comps, and measurement. The parts that solve it are swappable: agents, models, training methods, simulators and compute.
- Parity is enforced by the environment. Agents perceive only through the perception layer and act only through the agent interface, so no agent, however it works, can get around human limits.
- One agent interface. Every environment and every agent use the same observation and action format.
- Choices are data. A character's loadout (spec, talents, glyphs and gear) and its team's comp are inputs to agents, not built into them.
- Recordings are durable. Matches are recorded, full game state included, in a stable, versioned format that future methods can use.
- The server is authoritative. No client receives information its player shouldn't have.
Training and measurement#
- Agents improve by playing with and against other agents. Win rate on the real server is the final judge, never a simulator. Narrower drills and human insight are allowed, but a behaviour stays only if it raises rating. See measurement.
- The comp is the unit of training. A team trains together toward a shared win, and any agent in it can be replaced by a human.
- Measurement is automatic: agent-versus-agent ratings measure strength, and automated checks verify parity. Human play is informal feedback.