GPT-6 Astra: Moving from AI Chatbots to Computer Agents

How It Actually Works

The core shift in Astra is the distinction between Chat and Work. While the Chat mode is the traditional conversational interface we've used for years, the Work mode is designed for execution.

Chat vs. Work Mode

In Chat mode, you ask a question and get an answer. In Work mode, you assign a job. The AI can then use your files, connected applications, and your web browser to complete a multi-step process without you needing to prompt it at every single step.

As Nick Puru puts it, "Work is more or less built for people that are not technical that hides the code in the commands and it talks to you in very plain English." This allows non-developers to trigger complex workflows that previously required custom API integrations.

Plugins and Skills

To interact with the world, Astra uses two primary mechanisms: Plugins and Skills. There are currently over 1,400 plugins available, allowing the AI to connect to tools like Gmail, HubSpot, Dropbox, and Canva via OAuth.

Skills are essentially automated Standard Operating Procedures (SOPs). Instead of explaining how to write a proposal every time, you package those instructions into a 'Skill.' Once saved, you can simply tell the AI to "use the proposal skill," and it will follow the predefined steps perfectly.

Recording and Scheduling

One of the most powerful additions is the Record and Replay feature. Users can record their own screen while performing a manual task, and Astra converts that recording into a reusable skill. This effectively turns a human demonstration into a digital automation.

Additionally, Astra supports Scheduled Tasks. You can run a job once—such as pulling weekly booked calls from GoHighLevel—and then tell the AI to turn that into an automation that runs every Monday at 8 a.m. in the cloud.

Technical Deep Dive

From a technical perspective, GPT-6 Astra's 'computer use' capability represents a shift toward Agentic Workflows. Unlike traditional LLMs that generate a single response, an agent operates in a loop: Perception $ ightarrow$ Reasoning $ ightarrow$ Action $ ightarrow$ Observation.

The AI takes a screenshot of the current UI, analyzes the coordinates of buttons and text fields, and then sends a command to the OS (like a mouse click or keystroke). It then observes the new screen state to verify if the action worked before proceeding to the next step. This explains why OpenAI's computer use test score jumped to 73% from 66%, as the model has become significantly better at spatial reasoning and UI navigation.

The distinction between Local and Cloud execution is critical for the architecture. Local execution runs on the user's machine, allowing the AI to edit local files and control desktop apps, but it requires the computer to remain powered on. Cloud execution runs on OpenAI's infrastructure, meaning tasks continue even after the laptop is closed, though it cannot access local files unless they are explicitly attached.

For those who prefer more control, Astra includes a Codeex option. This is the same underlying model but exposes the code and commands being executed. This is similar to how 'Advanced Data Analysis' worked in previous versions, providing a transparent look at the Python scripts or API calls the agent is making in the background.

Why This Matters

The transition to computer-use agents solves the 'last mile' problem of AI. For a long time, the bottleneck wasn't the AI's intelligence, but the manual effort required to move that intelligence into a tool. If the AI wrote a great email, you still had to copy, paste, and send it.

By automating the 'busy work'—filling out forms, updating CRMs, and coordinating between apps like n8n, Beehive, and Slack—Astra reduces the need for complex middleware. We are seeing a move toward 'UI-based automation' where the AI interacts with the software exactly as a human would, bypassing the need for every single app to have a perfectly documented API.

As Nick Puru notes, "GPT6 Astra is the first AI that has truly felt like AGI to me" because it can handle a real job, walk away, and return with a finished product. This shifts the human role from 'doer' to 'manager,' where we provide the anchor (examples of quality) and the feedback, rather than performing the manual labor.