OpenAI introduced Dots at DevDay on September 29 as persistent agents with their own cloud computer and browser. They can connect to more than 4,000 applications and continue working on assigned projects in the background.
The launch came shortly after Meta introduced Muse, its personal AI agent built around a dedicated virtual machine. Google has also been moving Gemini in the same direction with Gemini Spark, a 24/7 agent that can continue working across Google services even when a device is locked or switched off.
These launches point to a broader change in how the biggest AI companies are designing assistants, moving from task-specific tools towards persistent, general-purpose agents.
From chatbots to persistent AI agents
The first wave of consumer AI was largely organised around individual tasks. AI then moved towards agents designed to complete specific tasks on a user’s behalf. The next shift is towards persistent, general-purpose AI agents that can handle multiple tasks over time.
OpenAI’s Dots and Meta’s Muse are among the clearest examples of this approach, while Google’s Gemini Spark is built around a similar idea.
OpenAI Dots bring the agent its own computer
OpenAI describes Dots as persistent agents that can continue working in the background, with each Dot getting its own cloud computer and browser.
This gives the agent an environment in which it can carry out tasks without relying on the user’s physical computer. Dots can browse the web, work with files and connect to external applications. OpenAI says Dots can connect to more than 4,000 applications.
The persistent nature of Dots is central to the model. A user can assign a project and the agent can continue working on it after the active ChatGPT session has ended. It can also retain context across ongoing projects, reducing the need for users to repeatedly explain the same task or provide the same information.
OpenAI is also building Dots around application access rather than keeping the agent within ChatGPT. Connected services can provide information and allow the agent to carry out work across different parts of a user’s digital workflow.
Meta’s Muse uses a dedicated virtual machine
Meta is taking a similar approach with Muse, its personal AI agent.
Muse operates through a dedicated virtual machine, which gives it a separate computing environment and browser. It can open websites, fill forms and interact with applications on the user’s behalf.
Meta says Muse can continue working after the user closes the application. For tasks that require a decision or approval, it can return to the user rather than completing the action on its own.
The company has also built controls around the virtual machine. Meta says its Sentinel system monitors Muse’s behaviour, while users can control which applications the agent can access and what permissions it receives.
Muse also retains information from previous interactions. Meta has described use cases where the agent can remember details about a user and use them in later tasks.
That memory is important for an agent intended to work across multiple activities. A task-specific agent may only need information related to the job it has been given. A persistent agent needs to maintain enough context to pick up work later and apply information from earlier interactions.
Google is moving Gemini in the same direction
Google has been developing a similar model through Gemini Spark, which the company describes as a 24/7 personal AI agent.
Spark can work in the background across Google services such as Gmail, Docs and Sheets. Because the agent operates in the cloud, Google says it can continue working even when a device is locked or switched off.
Google has described use cases where Spark can take information from Gmail, organise it in a Google Sheet and maintain documents or calendars based on ongoing instructions.
Spark also has controls for actions that could have consequences outside the AI system. Google says the agent asks for confirmation before high-stakes actions such as spending money or sending emails.
The cloud-based approach is similar to the architecture being used by OpenAI and Meta. Instead of requiring the user’s device to remain active, the agent has access to its own computing environment and can continue working independently.
Google is also changing how specialised AI works
The move towards general-purpose agents is also changing how users create specialised behaviour for AI.
Google has announced that its Gems system will transition to Skills from November. Gems allowed users to create customised versions of Gemini with specific instructions for particular tasks.
Skills are designed as reusable instructions that Gemini can recognise and apply when relevant. Google says Skills can also be combined to handle more complex tasks.
This changes the role of custom AI configurations. Instead of users deciding which individual Gem to open for each task, Skills can become part of the broader Gemini system and be applied when the agent determines that they are relevant.
Google says Skills can include reference material such as documents, PDFs and images. Existing Gems are expected to be migrated to Skills as part of the transition.
The cloud computer is becoming part of the agent
The emergence of persistent agents is also changing what is required to run consumer AI.
A conventional chatbot mainly needs a model, a conversation and access to the information provided in that interaction. An agent that works continuously needs somewhere to run software, access websites, store files, interact with applications and maintain its state between sessions.
That is why cloud computers and virtual machines are becoming part of the architecture.
OpenAI gives Dots their own cloud computer and browser. Meta uses a dedicated virtual machine for Muse. Google runs Spark in its cloud infrastructure.
The approach also allows agents to move between applications. Instead of a user manually opening email, a browser, a spreadsheet and other services, an agent can potentially move between those services while working on a single assignment.
This could make the distinction between individual AI features and a general-purpose assistant less important. The user does not necessarily need to know which application is being used for each part of a task if the agent can access the required services.
Persistent agents also raise new control questions
The move towards persistent agents also raises questions about how much autonomy they should have and what happens when they encounter situations they cannot handle.
AI agents can browse websites, fill forms, make bookings, send emails and complete purchases. But these tasks depend on systems and people outside the agent’s control. A website can change its layout, a product can go out of stock, or a customer service representative can respond in a way the agent was not trained to handle.
Meta’s Muse has already shown this limitation. The company has been testing a human concierge system for some phone-based tasks that Muse could not complete on its own. The system allows a human contractor to take over calls for tasks such as booking appointments and checking product availability.
Persistent agents create another problem when they move between websites and applications. An agent may have permission from a user to perform a task, but that does not automatically mean every website it visits recognises that authority.
This became clear after Amazon blocked Meta’s Muse from accessing its marketplace. Amazon said Muse was an unauthorised AI agent and raised concerns about how it identified itself and handled customer credentials. Meta disputed the security concerns and said Muse cannot see passwords or payment information, with credentials stored securely.
The issue is larger than one dispute. Most of the web was built around human users, with systems for authentication and access designed around a person logging in and carrying out an action. An AI agent acting on behalf of that person needs to establish who it represents, what the user has authorised it to do and what limits apply to that authority.
This becomes more important when an agent moves across several services. A user may authorise an agent to book a flight, but that permission could involve logging into an airline, making a payment and accepting terms on the user’s behalf. Each step can require a different level of access.
The challenge is to allow an agent to work for longer periods without turning its autonomy into unrestricted access. As agents take on more tasks and interact with more services, identity, permissions, human intervention and accountability will become part of how these systems operate.
Leave a comment