We use cookies to understand how you use our site and improve your experience. Cookie policy
Last updated:
A chat is one conversation with an agent inside a session. You can run several chats inside the same session, each on a different task.
A typical session has more than one chat going at once. For example:
All three chats work in the same sandbox. They can all read and write to the same files. But each chat has its own conversation history, so the agent doesn't get confused mixing tasks.
You see what each chat did when you review the session before saving back.
You pick the agent per chat. Inside one session, you can have one chat running Claude Code and another running Codex.
Why this matters:
You can also switch the model inside the same agent (e.g., a faster model for quick edits, a stronger model for deeper reasoning).
Type @ in a chat to point the agent at something specific: a file, a folder, or any added reference.
This is useful when:
The agent always has access to the whole workspace. @ is how you point at the part that matters right now. The same syntax works for references, the folders you've added from outside the workspace.
This is the key thing to understand:
So if one chat drafts a doc and another chat edits it, they're both working on the same file in the sandbox. But neither chat sees the other chat's conversation. Each is having its own conversation with the agent.
When you save the session back, all the changes from all the chats go together.
By default, each chat is sandboxed to the workspace folder. The agent can read and write inside the workspace, and that's it. It can't reach into your other files or system folders.
You can disable the sandbox per chat to give the agent full access to your computer. Useful when:
When the sandbox is off, you'll see the standard macOS permission prompts as the agent reaches into new folders (Desktop, Documents, etc.). You can approve or deny per folder.
Default: on. Turn it off when you need to, on a chat-by-chat basis.
By default, each edit the agent makes is shown so you can read it as it happens. Flip on Auto-accept edits to let the agent run without surfacing each one.
The changes still land in your session sandbox, and you still review them all before saving back. The toggle just removes the per-edit interruption while the agent is working.
Good for long-running tasks you trust. Off by default.
You don't have to wait for the agent to finish a reply before sending the next message. Anything you type while the agent is working gets queued and runs in order, as soon as the agent is free.
Useful when a follow-up idea hits you mid-thought, or when you want to line up several steps in a row.
Agents sometimes produce HTML: a landing page mockup, a prototype, a styled report. Any HTML file in your workspace previews in the right-side panel, the same way markdown files do. You can also open it in your browser.
This is one of the best ways to share AI-generated work with humans.
Start a new chat when:
Keep going in the same chat when: