Most of the work on AI in games goes into non-player characters: shopkeepers who answer in their own words, companions who remember what you said, townsfolk with goals of their own. That is what most people mean when they search for “AI NPCs”.
There is a second idea that gets less attention. Instead of making the characters around the player smarter, you give the player an agent that acts with the player’s own character, on the player’s goals, including when the player is not there. We call this an AI player agent, and it is what we build at Metashift Labs.
The two often use the same underlying models. They solve different problems, and they come with different design risks.
What is an AI NPC?
An AI NPC is a non-player character driven by a generative model rather than only by a script or a dialogue tree. Typical abilities:
- Dialogue: the character answers free-form questions instead of offering three fixed lines.
- Personality: a consistent voice, mood and backstory that shape how it replies.
- Reaction: it notices what the player does and responds, within the rules of the game.
The NPC still serves the game’s goals. It exists to make the world feel alive, deliver a story or give the player someone to deal with. Its success is measured by how believable and engaging it is.
A sourced example: at CES in January 2025, NVIDIA described “Smart Zoi” in inZOI, built with its ACE technology. In that description, every Zoi in the city acts on its own, driven by its own life goals. That is a generative AI character in the classic NPC sense: the world becomes more alive, and the player watches and interacts with it.
What is an AI player agent?
An AI player agent acts for the player. It controls or assists the player’s own character and works toward goals the player sets. Its success is measured by whether the player’s plan moved forward.
The same NVIDIA announcement from January 2025 points in this direction. NVIDIA presents ACE as moving toward autonomous characters that “perceive, plan, and act like human players”, and describes PUBG Ally as a co-playable character that can give strategic suggestions, find and share loot, drive vehicles and fight other players alongside you. Ally is a teammate rather than your own character, but it works on your goals in the match.
Research is moving the same way. Google DeepMind’s SIMA 2, described in a technical report published in late 2025, is a Gemini-based agent that follows instructions and acts inside many 3D games. The report shows task completion of roughly 45–75% in games it had not seen before, against 15–30% for the first SIMA, and notes it still struggles with long tasks and memory. It is a research agent, not a shipped game feature, but it shows that an agent can play on a person’s behalf across different worlds.
In Dominus Automa, our MMORPG, the player agent is the core of the game. Each player has a hero with an AI personality and moods. The player gives it a plan in natural language, by chat or voice, and the hero carries it out on our servers, around the clock, including while the player is offline. The hero reports back with numbers such as experience or pelts per hour and suggests changes. Talking to your hero through Telegram is in closed testing with more than 100 players. For more on the general idea, see our explainer on agentic AI in games.
AI NPC vs AI player agent: side by side
| AI NPC | AI player agent | |
|---|---|---|
| Whose goals it serves | The game’s: story, atmosphere, challenge | The player’s: a plan the player set |
| Who it acts for | Itself, as a character in the world | The player, through the player’s character or a teammate |
| When it runs | Mostly while the player is present and nearby | Can run while the player is away, if the game allows it |
| What success looks like | Believable, engaging interactions | The player’s progress, and a player who trusts the result |
| Main risks | Off-script or low-quality replies, cost per conversation, players’ doubts about generated content | Fairness of progress, cost of running agents for every player, players feeling the game plays itself |
Why player agents fit players who are short on time
Many of today’s players are adults with full schedules.
- The Entertainment Software Association’s 2026 Essential Facts, published in June 2026, puts the average US player at 37 years old.
- MIDiA Research data for Q4 2024 found that two-thirds of PC and console players play less than 10 hours a week.
- A YouGov survey from July 2024 asked gamers who now enjoy gaming less why. 47% said work or family responsibilities contributed “a lot”, and 45% cited having less free time.
Our CEO, Rafał Bieniek, put it this way on the Strefa Inwestorów Demo Day podcast in October 2026:
“Players aren’t losing interest in games. They’re losing the time to play them.” — Rafał Bieniek, CEO
An AI NPC does not help much with that. It makes the time you spend in the game richer, but you still have to be there. A player agent keeps your progress going when you cannot be.
Our August 2026 playtest on Steam is the evidence we have. Across the month-long playtest, more than 12,500 players played around 127,000 hours. About 60% of fights happened without the player present, and 43% of players used offline play, where the hero keeps going on the server after the player logs off. For the 7–23 August window, our own data shows 57% of play time was unattended. The full breakdown, including how the offline time limit changed behaviour, is in our write-up of the playtest data on heroes playing alone.
What are the design problems with AI player agents?
Agents that act for players raise harder questions than smarter NPCs do. Three matter most.
Fairness
In most online games, time spent equals power. If an agent plays while you sleep, the player with the most free time no longer wins by default, which is part of the point. But it also means the game must be designed for it from the start: shared rules for every hero, limits that apply to everyone, and game logic that runs on the server rather than on the player’s machine. Dominus Automa is server-authoritative for this reason. In our playtest, raising the offline limit changed how much of the game was played by heroes alone, so these limits are a design lever, not a detail.
Cost
An NPC talks when a player approaches it. A player agent may run for every player, all day. That changes the economics. In Dominus Automa a hero message costs a fraction of a cent today, our planned subscription is meant to cover server and AI model costs, and we plan to move to our own locally run models over time.
Player sentiment
Many players are wary of generative AI. Quantic Foundry’s December 2025 survey of 1,799 gamers, skewed toward core players, found that 85% view generative AI in games negatively, with older gamers somewhat less negative. Any game using AI has to earn trust.
Our approach is to use AI where it changes how the game plays, and to keep it out of what players see as craft:
“We keep AI in the process, not in the final product.” — Rafał Bieniek, CEO
In practice, the visible assets in Dominus Automa are made by artists. The AI lives in the hero’s behaviour and conversation, and in our development tools.
Will AI NPCs and AI player agents converge?
Partly. Both depend on the same building blocks: models that understand language, keep some memory and choose actions inside game rules. A companion like PUBG Ally already sits between the two: a character in the world that works on the player’s goals.
The difference that will remain is whose goals the AI serves. An NPC answers to the game’s designers. A player agent answers to the player. Games will likely use both: smarter worlds for players who are present, and agents for when they are not.
FAQ
Are AI NPCs and AI player agents built with the same technology?
Often, yes. Both can use large language models for dialogue and planning. The difference lies in what the AI is allowed to do and for whom: an NPC acts as part of the world, a player agent acts on the player’s instructions.
Does an AI player agent mean the game plays itself?
Not in the way we design it. The player sets the plan, reviews the hero’s reports and changes course. The agent handles the hours the player cannot spend, not the decisions.
Is Dominus Automa an AI NPC game?
No. Dominus Automa is built around player agents: each player’s own hero acts on that player’s plan, online or offline. Learn more in our overview of Dominus Automa as an automated MMORPG.