Agentic AI for MOBA (Multiplayer Online Battle Arena) game

Robot is an agentic live-ops crew for live-service games — a team of eight AI agents that each own one post-launch job end-to-end, running in production against our own real game. The game is a 3D browser-based 5v5 robot arena shooter (Three.js, live multiplayer, chat, and voice) whose match database records every shot, time-to-kill, and crash. Instead of humans staring at that telemetry, the agents read it: an impact agent statistically verifies whether each release actually helped; a balance copilot finds outlier weapons and proposes numeric fixes with citations, backtests, and 95% confidence intervals (validated on 3,347 real matches); BugForge turns raw player bug reports into runnable repro steps and clusters duplicates; a patch-notes generator writes notes from the code diff, not the commit message (100% change coverage, ~3.4s per release); and UA, creative-variant, and listings agents score ad creatives against pLTV ground truth (+5.7% ROAS in backtest) and localize store listings into 6 QA-checked locales. The design rule throughout: agents propose with evidence attached, humans accept — every proposal ships with a rollback path. Who it's for: game studios and live-ops/UA teams drowning in post-release toil — and, today, the players and operators of our own game, where the crew runs daily. Current product surface: everything is live, not slides. The playable game (AWS backend), an operator admin console hosting the balance copilot, impact reports, BugForge, UA and variants tabs, a public patch-notes site, and an Agora-powered voice support console where players talk to an agent that knows the game — all traced in Langfuse with continuous LLM-judge evals.