AI Features

AI Agent

A goal-driven mode of the AI Sidebar. The Agent runs multi-step sessions, edits across chapters, and waits for your approval on every change.

An Agent run mid-process, steps on the right, a pending proposal in the editor

How it works#

The AI Agent is the second mode of the AI Sidebar. Where Chat mode answers questions and never touches your text, Agent mode takes a goal, runs a sequence of steps to achieve it, and emits Proposals: concrete edits across one or more Chapters that you review and approve before they apply.

Three concepts are worth knowing up front:

Goal: a single instruction you give the Agent ("rewrite chapter 3 in the past tense", "draft an opening for chapter 5 from these sources", "tighten the dialogue across the final three chapters").
Step: what the Agent did at each stage of the run. Steps include creating chapters, renaming, moving, reading sources, and producing proposals. The list is visible in the sidebar as the run progresses.
Proposal: a structured suggestion targeting one Chapter. It contains one or more edits (replace, insert before, insert after, delete) and shows up as a diff in the Editor for you to approve.

Switch to Agent mode#

In the AI Sidebar, click Agent at the top of the input area to switch modes. The pre-fill prompts disappear (the Agent expects a goal instead of a conversational question). The Tools menu (Web search, Image generation) is unavailable in Agent mode.

To switch back, click Chat.

Start a run#

Open the AI Sidebar and switch to Agent mode.
Optionally pin the chapters and Sources you want the Agent to focus on, via the context menu above the input.
Type your goal in the input and press Enter.
Watch the run unfold in the sidebar. Each step appears as it completes. Proposals arrive inline.

The Agent runs against the Chapter you're currently viewing (it's tied to a chapterId) plus any pinned context. To target a different chapter, click it in the Chapter Tree before starting the run.

Review a proposal#

When the Agent generates a proposal, the affected Chapter shows a red/green diff in the Editor:

Red strikethrough: text the Agent wants to remove or replace.
Green: the Agent's proposed new text.

A bar at the top of the Editor lets you navigate the proposal's edits one by one:

Accept: keep this edit.
Reject: discard this edit.
Accept all: keep every edit in the proposal.
Reject all: discard every edit and ask the Agent to revise.

You can accept some edits and reject others within the same proposal.

Continue a run#

After the Agent finishes a step, you can keep the session going by typing another message in the sidebar, for example to refine the goal or push back on a rejected proposal. The run extends rather than starting over, so context is preserved.

Cancel a run#

While the Agent is working, click Stop in the sidebar to abort the run. Partial steps and any pending proposals remain visible; you can accept what's already been emitted or discard it.

Resume a past run#

Click History at the top of the sidebar. Past Agent runs appear in the same list as Chat conversations, newest first, with the Agent's goal as the title.

Click an item to load it back into the sidebar.
Send a new message in the sidebar to continue the run from where it left off.

Limitations#

One run is active at a time per Project. Starting a new run requires the current one to be stopped or completed.
The Agent can create, rename, move, and edit Chapters, but cannot reorder them in the Chapter Tree.
Tools available in Chat mode (Web search, Image generation) don't apply to Agent runs.