Reviewing and controlling agent changes | Ep 3 of 6 - VS Code Learn

Episode 3 — Reviewing and controlling agent changes

Scope

This episode demonstrates how to steer, edit, review, fork, and restore work done by a VS Code “agent” (automated code assistant) while it edits files and runs tasks in your workspace. It provides a practical walkthrough covering:

• Changing a prompt and steering agent behavior mid-action
• Reviewing diffs and running generated code/tests
• Branching experiments via session forks
• Reverting unwanted changes via checkpoints

---

Key product features & UI elements shown

Edit a message / prompt

• Click a message to modify the agent’s instructions. This keeps a cleaner action history compared with many follow-up messages.

Steering agent actions

Options are available from a drop-down beside a message:

• Steer with message — sent at the agent’s next yield opportunity.
• Stop and send — stop the current action and send a message immediately.
• Add to queue — enqueue a message to run after the current request completes.

Change review UI

• After edits, a status/summary bar above the chat shows files changed and lines added/removed (example: “changes to 2 files, 78 lines added, 2 removed”).
• Options to accept changes:
  • Keep all changes.
  • Keep per profile.
  • Open files and accept/reject individual changes (additions shown in green, removals in red).
• Ability to inspect and run generated code/examples immediately to verify behavior.

Session branching and checkpoints

• Fork session: use /fork or the fork icon to create a new session that inherits history — useful for exploring alternate implementations (for example, keep a CLI session while experimenting with a FastAPI refactor in the fork).
• Restore checkpoint: revert the workspace to a prior checkpoint to undo agent edits or the last request — useful when experiments produce unwanted changes.

Agent-run tasks and terminal

• Agents can execute tests and commands.
• You can view the agent’s hidden terminal and see test run output (for example, when an agent converted a CLI to FastAPI and updated tests).

Workflow tips

• Edit prompts to request specific changes (e.g., “input should be provided by a CLI argument”).
• Use steering to add follow-up tasks like “also add tests” while the agent is working.
• Use forks to parallelize exploration and checkpoints to safely experiment and roll back.

---

Tutorial / guide elements covered

1. How to edit an agent prompt/message to change behavior.
2. How to steer an agent mid-action (steer, stop & send, add to queue).
3. How to review agent-made changes (summary, diffs, accept/reject at file or hunk level).
4. How to run and verify generated code and tests.
5. How to fork a session to explore alternate directions without losing the original.
6. How to restore to a previous checkpoint to undo unwanted edits.

---

Takeaway / next steps

This episode covers practical control and review tools for safely iterating with agents: editing prompts, steering actions, accepting or undoing changes, forking sessions, and reverting via checkpoints. The next video will cover the session view for monitoring multiple sessions and ongoing work.

---

Main speaker / source

VS Code Learn instructor / demonstrator (narrator of the VS Code Learn episode).