Battle Drones: simulated balance, player fun and a cheap-to-host game

Brief and result, October 21, 2026. Question: when an agent implements and simulates a game, which decisions still need a player and what makes the release maintainable? A useful answer tracks prompts, simulations, human play, code review, network/deployment changes and later use. Our Fallout: New York case shows a first playthrough can overturn a generated-world acceptance test. This case has richer public prompt/version and repair evidence but remains owner-built, not independently requested software.

Brief's three subquestions: (1) What do the prompts, versions and repository prove about human design input? (2) What did simulated games test versus a person's sense of fun? (3) What happened to sound multiplayer and the proposed static-hosting economics after actual play? Checked Willison's October 20 first-person history and updated post, project repository, online multiplayer change and audio warning change; searched for independent repeat-player evidence and found none in these materials. Follow up on independent play only if a public repeat-use or repair trace appears.

Need→design→candidate. The first mobile real-time-strategy prototype began in July 2026; by October Willison could play it but found it dull. The agent drew new sprites and maps, yet the owner's later judgment still found it boring. He gave an open-ended fun brief; Claude's overnight v4 added match openers, contested objectives, comeback mechanics, modes, feedback and progression. The agent-generated release notes report 1,200 AI-vs-AI normal-difficulty matches, fewer drawn-out matches and more wins by resource-disadvantaged teams. These measure simulated match duration and comeback frequency, not whether players enjoy or return to a game; author says from his own play that v4 was fun. The article does not provide a controlled player trial or active-human-hour/accounted-token total. Its count of 321 agent commits by the writing of the post is a change count, not production value.

Human product decisions persisted. After v7 added infantry transport, Willison noticed the AI seldom used the units. For v8 he specified islands whose objectives require infantry flown there, plus bridges, limited resources and a replacement transport so the strategy remained available. His prompt asks the agent to research and critique before implementing an agreed design. For the giant mobile base he brainstormed damage and transport rules with the agent. He says he reviewed almost none of the TypeScript and barely knows the language; human play and explicit constraints here are product/design steering, not code assurance. The initial release notes' match data cannot attribute fun to a single mechanic, and player judgments cannot establish stable balance.

Real use changed both safety and hosting. The October 20 article initially said audio transport lets two nearby devices play while keeping the game static and avoiding a server Willison would have to operate; he reported quick tests on laptops, phones and tablets. In the article's later update, playing with Natalie caused a headache from the high-pitched modem sounds. The audio warning change discloses the sound risk rather than proving it harmless. Willison then tried WebRTC without success between his own devices and had Claude build a stateless WebSocket relay; the October 20 merged PR contains the client, small Python server, deployment guidance and tests for its protocol. Online play now works by sharing a room link or QR code according to the owner. He estimates an always-hot shared Fly instance at up to $2.19/month, not an observed bill or a guarantee of scalable zero-maintenance service. The relay restores an ongoing deployment, availability and abuse surface the sound-only approach was meant to avoid. Browser determinism also needed custom trigonometry after Chrome/Safari differences desynchronized replay, per the author. It is an informative same-feature repair prompted by a person, not customer acceptance after a month.

October 21 follow-on: Willison links a French-language steampunk reskin named Panache, made by another person a day after he released Battle Drones under Apache 2.0. It reportedly changes artwork, music and an animated opening more than gameplay. This is evidence of an external creator adapting the released game, not independent repeated play, a production maintenance contract, or a documented coding-agent workflow in the fork. Attribution of the fork to an agent is not established by the short source.

Limits and next measurement. Public owner logs and merged repository artifact show prompt→iterative design→player harm report→warning→network alternative. They do not show an external requester's acceptance, independent player retention, headache incidence, accessibility test, review of generated code, active human design/review/repair minutes, model bill, server bill or later multiplayer uptime. Compare game modes across human playtests with time-to-first-fun and return sessions alongside simulated win/balance measures; record sound/device adverse effects and time from report to patch, deployment and player retest. For economics count agent runs, human balancing/review, ongoing server and moderation/support; avoid claiming a static artifact has no maintenance burden. Interface versus service now has an example in which one harmed player prompted a change of runtime architecture. Roles likewise includes a human who both judged fun and rejected an unsafe interface.

Backlog: Willison's firsthand article is long (~15-minute operator walkthrough). Hold for Saturday October 24 if this thread earns feed room or Dru opens one of the existing cards; then lead with player-reported headache forcing the static game to acquire an online server, not the 321 commits. A shorter standalone note on the reversal could work on a weekday after backlog clears. The eMudhra certificate item received the same subscription night and is security/package news outside this area's question; do not recast it as agentic software evidence.