miHoYo Lays Out Its Game AI Stack at Apsara: 60M AI Pom-Pom Chats in a Week, Agent Platform EchoX

miHoYo used the closing Apsara Conference to lay out something close to a full roadmap for AI in games. Zheng Yinhe, who leads AI NPCs and gameplay for the company's Honkai series, described a loop the studio sums up as AI entering the game and the game feeding AI back: make characters feel human first, let them act inside gameplay, then send AI into the production pipeline and use the game itself as a training ground.

The clearest numbers came from Pom-Pom, the train conductor in Honkai: Star Rail. miHoYo gave the character AI dialogue in April, and within one week it drew more than 60 million conversations — one player talked to it 1,379 times in a single day. Offered as a beta called Pom-Pom Bang Bang, the feature ran for just over a month before being taken offline for optimization.

Three pieces of engineering produce the human feel. The system tags emotion alongside generated text and drives a 3D model to produce matching expressions and gestures; a strategy-enhanced RAG layer sits on top of ordinary retrieval from official and community sources so answers pass through the character's own view of people and its current mood; and memory is split into short-, medium- and long-term tiers that hold player preferences and past events.

Zheng framed the hard problems as a set of engineering gates: keeping reply quality stable over time, keeping up with a game that updates every 42 days, and serving tens of millions of players at once while controlling cost. The core challenge for AI Pom-Pom was making a character speak reliably.

AI Gameplay pushes further. The model has to talk in character and also decide the next action from what the player said and the state of the board, which makes conversation itself part of the game process. Zheng also teased a new attempt in that direction: a board game where AI characters judge the situation, compete, and remember grudges built up with the player, used to test whether a character can hold its persona while making real choices.

The half players never see lives on the production side. EchoX, an internal platform, hosts code agents and a matching tool ecosystem, with a Harness layer defining how agents work and MCP connections wiring in logs, the engine and internal tools. In one demo an agent read logs, located a performance bottleneck and applied the fix; in another, a designer handed over a gameplay request plus a UI sketch and got back a playable white-box demo. Material swaps, three-view extraction from concept art and movement generated from dialogue are also handed to AI.

Cost is the other gate. Honkai: Star Rail producer Jiang Dawei said earlier at a campus talk that dozens of agents collaborating for 13 hours burned 2 million yuan worth of tokens. The bill came from a multi-agent experiment unrelated to any specific gameplay, but it points at the second problem in agent-driven development. Getting agents to work is one hurdle, keeping them from idling, drifting and burning money is another.

The studio also treats the game itself as a training environment. It started with GUI agents built on Qwen-VL models fed more than 10,000 hours of gameplay footage and matching inputs, generating keyboard and mouse commands from the screen at close to 30fps. Frame-by-frame reaction ran into trouble on tasks that need long-term memory and planning, so the team switched to having a coding agent write a strategy script, run it, read the logs, revise and rerun. In an experiment on Balatro, the agent found an online simulator and started peeking at future card decks — a reward-hacking failure inside what amounts to a recursive self-improvement loop.

miHoYo's stated ceiling for AI spending is up to 100 billion yuan over three years, a bet co-founder Dawei Ge has described as setting off one big firework even if it fails. At a recent campus recruitment session at Shanghai Jiao Tong University, he said progress has accelerated since the company began building coding models this year, that he is confident miHoYo will be a significant player among Chinese large-model teams within two to three years, and that people can come slap his face in a year or two if that does not happen.