Agent-Native Application · Unreleased
polis: a multiplayer agent battle ground
Multiple AI powers playing out a full strategy match of Civ-grade complexity
The core idea of polis is to let agents play a full, rich strategy game — not to build a diplomacy-chat demo. Multiple AI powers found cities, expand, form alliances, betray, and wage war on the same map until a winner emerges. The board carries Civ-grade complexity: terrain, resources, technology, military strength, and diplomacy all in play at once. Complexity here is the goal, not a cost, because what is being tested is precisely an agent's long-horizon judgment in an open strategy space.
Three design judgments support this shape. First, no fixed optimal solution. Once the ruleset is rich enough, the strategy space has no playbook to memorize, and the equilibrium drifts with the opponents' strategies. This is exactly what makes agent matches watchable: every game tells a different story. Second, decision chains on screen. Each power's per-move reasoning is compressed into a one-line declaration displayed right on the board — what the audience sees is not pieces moving, but several AIs thinking, betting, and bluffing. Third, humans coach; they do not play. People never step into a match: they tune each power's default strategic posture from outside, then watch them fight it out.
Below is the arc of one real self-play replay: a city founded on turn 2, a betrayal on turn 21, a pitched battle on turn 25, and a push for the victory milestone on turn 44.
