The video demonstrates how to effectively manage and control agent-driven code changes in VS Code by editing prompts, steering agent actions in real-time, reviewing modifications, forking sessions for parallel experimentation, and restoring checkpoints to undo unwanted changes. These features enable a streamlined, flexible workflow that ensures code quality and supports dynamic development within the same project environment.
In this video, the focus is on managing and controlling the changes made by an agent within VS Code. Building on the previous video’s discussion about approval levels for agent commands, this session demonstrates how to steer ongoing work and review modifications effectively. Instead of sending multiple follow-up messages, the presenter prefers editing existing messages to maintain a cleaner workflow. By editing a message to specify that input should be provided via a CLI argument, the agent removes the previous request and updates the code accordingly, showcasing a more streamlined interaction.
The video then explores how to steer the agent’s actions during execution. Using the interface’s dropdown options, users can send messages to steer the agent, stop the current action, or queue messages for later. This allows for dynamic control, such as requesting the addition of tests while the agent is still working. The presenter highlights the ability to view generated code in real-time and demonstrates how the agent can respond to steering commands by implementing requested features like tests and CLI argument handling.
Next, the video covers reviewing the agent’s changes. After the agent finishes editing files, a summary appears showing the number of files changed and lines added or removed. Users can review changes file by file, with additions highlighted in green and deletions in red, and decide whether to keep all changes, keep changes per file, or selectively accept individual modifications. The presenter tests the updated CLI functionality to confirm that the changes work as intended, reinforcing the importance of reviewing and validating agent-generated code.
The concept of forking sessions is introduced as a way to explore different development directions without losing progress in the original session. Forking creates a new session with the entire history carried over, allowing parallel experimentation—such as refactoring a CLI to a FastAPI—while preserving the original session for other tasks like adding examples or tests. This feature supports flexible workflows and experimentation within the same project environment.
Finally, the video explains the restore checkpoint functionality, which allows users to revert the workspace to a previous state by undoing recent changes made by the agent. This is particularly useful when experimenting or when unwanted large changes occur. The presenter demonstrates restoring to a checkpoint after a significant refactor, emphasizing how this feature safeguards against undesired modifications. The video concludes by summarizing the covered topics—editing prompts, steering agents, reviewing changes, forking sessions, and restoring checkpoints—and previews the next video’s focus on managing multiple sessions.