For most of the last few years, “AI in games” meant conversation: a character you could type to, and that would answer in a voice that sounded less scripted. You still had to be there, at the keyboard, for anything to happen.
The next step is AI that acts. Instead of only replying, a character takes a goal, works out how to reach it, does the work inside the game and tells you what happened. This article explains what agentic AI in games means, where it already shows up, how we use it in Dominus Automa, and which problems nobody has fully solved yet.
What is agentic AI in games?
An AI agent in a game is a character driven by a loop, not a single reply:
- Goal. Someone, usually the player, says what should happen: “level up in the forest”, “gather pelts”, “don’t fight anything above my level”.
- Plan. The agent breaks the goal into steps it can carry out in the game’s rules.
- Act. It moves, fights, gathers and trades, using the same actions a player would.
- Report. It tells the player what it did, how well it went, and what it needs decided next.
A chatbot does only part of this. It responds when spoken to and stops when the conversation ends. An agent keeps working toward the goal after the conversation is over, and its actions change the game state. That is the difference between a character that talks about killing the wolf and one that goes and does it.
Autonomous game characters are not new as an idea: scripted companions and bots are a familiar part of games. What has changed is that language models let a player give instructions in plain words, and let the character explain itself in plain words, instead of through menus and fixed behaviour trees.
Where AI agents in games already appear
Several large companies have shown agent-style systems. They aim at quite different things, and it helps to keep them apart.
Google DeepMind SIMA 2: a research agent that plays 3D games
Google DeepMind’s SIMA 2 (November 2025) is a Gemini-based agent that acts inside 3D games from instructions. In games it had not seen before, DeepMind reports task completion of roughly 45–75%, against 15–30% for the first SIMA. The same report is open about the limits: SIMA 2 is still weak at long tasks and at memory. It is a research system for general agents, not a feature shipped in a game.
NVIDIA ACE autonomous characters
At CES in January 2025, NVIDIA presented ACE autonomous characters in several games: PUBG Ally, an AI co-player that fights alongside you in PUBG; Smart Zoi in inZOI, where characters act on their own in a life simulation; and AI teammates in NARAKA: BLADEPOINT that join your party. These are agents beside the player: teammates and characters that react to what is happening around them.
Microsoft Muse: a model of gameplay, not a player
Microsoft Research’s Muse (WHAM, published in Nature in February 2025) is a generative model that produces gameplay sequences to help with ideation. It is often grouped with game agents, but it is not one: it does not play the game for anyone. It is a tool for the people making the game.
Inworld: from NPCs to a general runtime
Inworld started out building AI for game NPCs. In August 2025 it introduced Inworld Runtime, positioned as a general AI runtime for consumer apps rather than only for game characters.
Agents for the player, not just beside the player
Most of the systems above put AI next to the player: a smarter teammate, a livelier town. We build AI that plays for the player. In Dominus Automa, our automated MMORPG, each player has a hero, and the hero is the agent.
How it works today:
- The player sets the plan. The player says what they want, in text or by voice, through in-game chat. Player-set hero paths are part of the plan for the November 2026 playtest.
- The hero acts around the clock. It plays on our servers, inside the same shared world as everyone else, whether or not its player is logged in. The game is server-authoritative, so the hero follows the same rules as any other character.
- The hero reports back. It reports how efficient it is at a location, for example experience or pelts per hour, and suggests changes. It asks for decisions and remembers the answers.
- The hero has a personality. It speaks in natural language and has moods, so a report reads like a message from a character rather than a log line.
In closed testing, more than 100 players can also talk to their hero through Telegram, so the hero can reach them outside the game. We plan to bring Telegram to the November 2026 playtest; it is not part of the public game yet.
We also run two experiments that are not features: a nightly newspaper written by AI from the server’s logs, and a “Rogue AI” idea built around a cursed artifact. Both are being tested to learn what players respond to, and may not ship.
The reason for this direction is time. As our CEO Rafał Bieniek put it in a Strefa Inwestorów Demo Day interview (15 October 2026):
“Players aren’t losing interest in games. They’re losing the time to play them.”
The August 2026 playtest is the first real test of that idea. Across the month, about 60% of fights happened without the player, and 43% of players used offline play. We break those numbers down in what happened when heroes played alone in our playtest. For how this approach differs from smarter non-player characters, see AI NPCs vs AI player agents.
What is still hard
An AI that plays for you sounds simple. Building one for thousands of players in a shared world raises problems that the whole field is still working on.
Long tasks and memory
DeepMind’s own SIMA 2 report (November 2025) names long tasks and memory as weak points. A hero that plays for hours has the same problem in a game setting: it has to keep a goal in mind across many sessions, remember what the player decided last week, and not repeat a mistake it already reported. Giving the hero a plan to follow, and recording the player’s answers, helps; it does not make the problem go away.
Cost per action
Every message from a language model costs money, and a hero that acts all day can produce many. In the same interview, Rafał Bieniek said a hero’s message costs “a fraction of a cent”, which is our own figure, not an independent benchmark. The design rule that follows is to use the model where language matters (instructions, reports, decisions) and keep the moment-to-moment game logic on the server, where it is cheap and predictable.
Fairness
If a hero plays while its owner sleeps, does more time online still mean more power? In a shared world, an agent changes the balance between players who can play for hours and those who cannot. That can be the point: a player with little time is no longer left behind. It also means progression, limits and rewards have to be designed with agents in mind from the start, not added later.
What players think of generative AI
Player sentiment is the hardest problem of all. A Quantic Foundry survey (18 December 2025, 1,799 respondents, skewed toward core gamers) found that 85% view generative AI in games negatively, with older gamers less negative than younger ones. Developers are wary too: in the GDC 2026 State of the Game Industry report (January 2026), 52% of respondents said generative AI is harming the industry, while 36% said they use it at work.
An agent that saves a player time is a different offer from AI-generated art or writing, but players will not separate the two automatically. The burden is on the game to show what the AI does for the player.
What to watch next
- Agents that last longer. The research gap DeepMind names, long tasks and memory, is the one that matters most for heroes that play for days.
- Agents that reach the player outside the game. Messaging channels such as Telegram turn a game into something you can check on, not only something you sit down for.
- Real usage, not demos. Showcases prove a system can work. Playtests with thousands of players show whether people use it. Our August 2026 playtest data is one public example.
If you are looking at this space from the business side, our article on how to judge AI gaming startups covers which signals to look for.
FAQ
What is the difference between agentic AI and a game chatbot?
A chatbot answers when you talk to it. An agent takes a goal, plans, acts in the game and reports back, and it keeps working after the conversation ends.
Is there AI that plays games for you?
Yes. Research agents such as Google DeepMind’s SIMA 2 play 3D games from instructions. In Dominus Automa, each player’s hero plays on the server while the player is away, following the player’s plan.
Are autonomous game characters the same as AI NPCs?
No. AI NPCs make the world’s characters smarter. A player agent acts for the player’s own character. We explain the difference in AI NPCs vs AI player agents.