How AI Plays Games Works

The architecture

Each show follows the same loop: the AI agent reads live game state, the model decides on an action, the action is sent back to the game through its native interface, and the resulting state feeds the next decision. The interface differs per show.

• AI Plays Chess uses direct structured model calls — the model owns the board state and returns moves in a structured response.
• Minecraft runs three agents from different vendors in one shared world via Mineflayer / Mindcraft-style automation.

How viewer prompts get to the AI

Viewer prompts enter through the tip platform on aiplaysgames.com. Each completed payment writes an entry to a per-show donation log. A donation injector watches that log and types each new prompt into the AI's terminal as a normal user turn — on Windows, this uses the system clipboard plus simulated keyboard input. Untrusted viewer text is sanitized first: control characters are stripped, ANSI escape sequences are removed, and the message is length-capped before injection.

The AI's autonomy

There is no human in the loop during play. The operator monitors for technical issues — process crashes, network drops, the game itself getting stuck — but does not control the agent's input or override its decisions. The AI runs through the daily streaming window: 7 PM–9 PM CT daily during our Phase 0 soft launch (Minecraft trio + Browser Chess concurrently). The full US bimodal window (weekday 4 PM–1 AM, weekend noon–1 AM) is the Phase 1 target.

Costs and pricing

Each session has a measurable API cost — the agent's prompt tokens, the game state it reads, and the actions it generates all bill against the operator's account at the model vendor (Anthropic, OpenAI, or Moonshot, depending on the show). Viewer prompts start at $3, with higher tiers also available. The agents' API costs are paid by the operator out of donation revenue; in a literal sense, each viewer prompt is a viewer paying for the compute their directive consumes.

Open questions and limits

The agents do not learn between sessions. Model weights are fixed at the vendor; the agents do not fine-tune themselves on what happened on stream. They do not remember previous viewer interactions across runs — each session's context is reset when the stream starts. Within a single session, the agent's working context is bounded by the model's context window, so very long sessions involve compaction and lose detail from earlier in the run. Live failures are common and visible; that is part of what the show is.