Action model
An action model is an AI model that returns the next action to take in an app or on a screen, such as a tap or a keystroke, where a language model would return text. AGI, Inc. builds AGI-ACT, a family of action models that complete tasks in the apps on a phone.
Last updated
Related terms
On-device action model
An on-device action model is an action model small enough to run on a phone's or PC's own chip.
Small action model (SAM)
A small action model (SAM) is an action model compact enough to run on a phone, PC or other edge device.
Computer use
Computer use is an AI agent operating a computer's apps and websites through the screen, the way a person does.
Vision-language-action (VLA) model
A vision-language-action (VLA) model takes an image and an instruction and returns an action, for a robot or on a screen.
Returns | The next step in an app's interface |
|---|---|
Works in | Apps with or without an API |
AGI product | AGI-ACT, from 2B to 20B+ parameters |
Research result | AGI-0: 100% on AndroidWorld (116/116 tasks) |
How is an action model different from a language model?
A language model answers with text, and a person still has to do the task. An action model does the task in the app, through the same screens a person uses, so it also works in apps that have no API.
Model type | What it returns | Needs an API in the app | Example task |
|---|---|---|---|
Language model | Text | Not applicable | Summarize a message thread |
Function-calling model | A call to a function the app exposes | Yes | Add an event through a calendar API |
Action model | The next step in the app's interface | No | Change a setting several screens deep |
What can an action model do today?
Action models handle multi-step tasks in everyday apps. AGI's agent, built on AGI-ACT, completes tasks like these on Android:
Ordering groceries and meals
Booking rides
Sending messages
Changing settings and filling out forms
How are action models measured?
Public benchmarks give an action model real tasks and check the result. On AndroidWorld, 116 tasks in 20 real Android apps, AGI-0 completed all 116, a 100% score on the official leaderboard. On OSWorld, which tests desktop computers, AGI's agent scored 76.26% in October 2025.
FAQ
Is an action model the same as an AI agent?
No. The action model is the AI model that picks each next step. The agent is the product built around it, which carries a request across apps and asks the person before sensitive steps such as payments.
Does an action model need an API for each app?
No. An action model works through the app's interface, the same screens a person uses, so it can complete tasks in apps that never built an API or an integration.
What is AGI-ACT?
AGI-ACT is the family of action models built by AGI, Inc., sized from 2B to 20B+ parameters. The models run on the device through AGI-RUN, AGI's inference engine, or in a hybrid setup with cloud or on-prem components.
