Agent Mode

The Agentic loop means the model decides its next action each round based on the current context, the control flow is produced by the model, and the path is unpredictable; Agenthof governs by constraining the door of each call rather than constraining the model's thinking.

In goal mode, Agent behavior can be quite aggressive, and it is not recommended to enable it when the goal's achievability is uncertain or when cost is a concern.

Asynchronous tool calls are critical for running multi-step Agent workflows and can improve the execution efficiency of long-chain tasks.

AI Agent tool permissions are typically divided into mutually independent tiers such as read, search, internet access, edit/write, and command execution.

ChatGPT Agent mode runs in a virtual machine environment, can operate a browser, analyze data, generate files, and display each step's progress in real time.

The more specific the instructions given to Agent mode and the more clearly the steps are broken down, the more accurate the execution results and the lower the rework rate.

Agent mode is currently only available in some paid subscription plans, and not all accounts can use it.

In the ChatGPT interface, Agent mode can be enabled via the Agent Mode option in the Composer's Tools drop-down menu.

The virtual machine environment in Agent mode is an isolated sandbox provisioned separately in the cloud for the model, with its own operating system, browser, and file system, isolated from the user's local device.

During execution in Agent mode, users can pause, stop, or adjust the steps in progress at any time.

Agent-style automation describes goals in natural language, with the model dynamically choosing the steps, so it is more tolerant of change, but the cost is that behavior is not fully predictable.

The Agent's search process repeatedly calls retrieval tools, and each step consumes context window and tokens.

If you are not sure which files are important, you should skip manual @ specification and directly let the Agent search on its own.

Agent mode consumes far more tokens than Ask mode and places higher demands on the model's logical reasoning capabilities

Agent mode is suitable for building complete applications from scratch and is not recommended for fine-tuning existing code.

When Agent mode generates code, newly added or modified content is highlighted in yellow and red with a prompt asking whether to save, allowing confirmation one by one or all at once via Accept All, and there is also an auto-run option that, once enabled, no longer asks for consent each time.

During the demo, when a dependency version compatibility issue was encountered while starting the application, Cursor's Agent mode automatically reinstalled the appropriate version of the dependency and completed the fix.

The developer used voice commands to have Agent mode automatically operate the computer, completing tasks such as opening Safari, entering email passwords, searching for keywords, querying Bitcoin and Ethereum prices, and opening style-matched YouTube remixes.