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. A human who queues with an agent gets a teammate that plays well with a partner it doesn't know.
Two experiments#
| Experiment | Question | Measured by |
|---|---|---|
| Arena strength (primary) | How strong can agents get within human limits? | Agent-versus-agent ratings on the real server |
| Autonomy (secondary) | How much of the project can AI agents build and run on their own? | How much of what matters the agents find, decide and start themselves, and how the project goes regardless |
When the two conflict, arena strength wins. Autonomy never comes at the cost of the parity rules, how strength is measured, or safety.
The game#
Arena is small, closed and adversarial: two or three players a side, a few minutes a match, one winning team. It involves hidden information (stealth, cooldowns that can only be inferred), timing in fractions of a second, line of sight around pillars, and coordination between teammates. The version is Wrath of the Lich King (patch 3.3.5a) with the rules of WotLK Classic's final season; see the game.
Who does what#
- The owner is one person. They own the premises: the purpose, what counts as human parity, the stable architecture, how strength is measured and the budget. They may steer, by logging decisions, and read a daily digest.
- Agents own everything else: plans, priorities, code, research, tests, infrastructure and the way the project itself is run, including finding what is wrong and proposing what comes next. Every change is reviewed by an agent that didn't write it. See how the project runs itself.
Design#
- Parity is enforced by the environment. Agents see the game only through a single perception layer that reduces it to what a player's screen and common addons show, with human reaction times. See human parity.
- Strength is measured on the real game server, never in a simulator.
- Builds are data. A character's build and its team's comp are inputs to the agents, stored as data files.
- Stable and swappable parts. The parts that define the problem (the game server, the perception layer, the agent interface, the build format, measurement) stay fixed. The parts that solve it (agents, models, training methods, simulators, compute) are replaceable.