Using the Agentic Brain

The Agentic Brain is the connected understanding the app builds about one AI opportunity. It uses the Canvas, approved sources, conversation history, computed assessments, and open questions to help the team improve that opportunity over time.

Instant Chat and the saved Brain are different surfaces

The front-page Instant Chat is an intake and preview experience. Canvas AI gathers the initial business context, may research a public website, and seeds a first Canvas, systems map, readiness assessment, remediation worklist, and directional benchmark. It does not expose the Chat, Guided, and Agent dial.

After the Canvas is saved, the Agentic Brain page can work across the whole Canvas. Section chat uses the same Brain but focuses on the active Canvas section. AI Readiness, Agentic ROI, Overview, Summary, Slides, and Outputs use read-only Brain conversations so an explanatory question cannot silently change the model.

Chat, Guided, and Agent modes

The mode controls who approves application. Chat means the user drives every change; Guided means the Brain proposes and always asks before applying; Agent means the Brain applies autonomously within visible bounds. (Before Spec 045, these modes were labeled Manual and Auto; internal identifiers are unchanged.)

Chat

Chat is the default and most review-heavy mode. It was previously labeled Manual.

• Asking a question produces an answer, recommendation, or proposed refinement.
• A chat response alone does not update Canvas content.
• When an Apply action is available, the user reviews the proposed destination and chooses whether to apply it.
• In a Canvas section, quick actions such as Suggest, Review, Shorten, Expand, Rephrase, and Format remain user-directed.
• Chat is the best choice for sensitive language, uncertain evidence, or any change that should be considered individually.

Guided on the Whole Canvas

Whole-Canvas Guided is a review-then-apply batch for open Brain gaps.

1. The user chooses Plan all gaps.
2. Brain checks the current open gaps, prioritizing weaker readiness areas first.
3. It prepares proposed section or workflow fixes and shows a checklist. Nothing has been applied at this point.
4. The user can cancel or choose Approve & apply for the batch.
5. Approved items are applied sequentially. The checklist reports completed, skipped, and failed items, and the user can stop the remaining work while it is applying.

One batch approval does not authorize unrelated work. Guided stays inside the current Canvas and the listed open gaps.

Guided in a Canvas section

Section Guided follows the same rule as the Whole Canvas: Guided always asks before applying. Selecting Run Guided has Brain review the active section with Canvas context and prepare the single highest-impact improvement — as a proposal. The proposed change appears in the chat as a Proposed Refinement card, and the section draft is untouched until the user chooses Apply Action. (Before Spec 045, section Auto applied its one update immediately on click; that behavior is superseded — see Spec 004 DEC-017.)

Agent on the Whole Canvas

Whole-Canvas Agent is a visible, bounded gap-resolution loop.

1. The user chooses Run agent, authorizing work on the currently listed open gaps.
2. Brain shows a live plan and takes one gap at a time.
3. Each successful step updates the mapped Canvas content and resolves the corresponding Brain gap.
4. The loop ends when no open gaps remain or when the user chooses Stop. Errors and skipped steps remain visible.

Agent mode does not deploy production agents, contact external people, or continue outside the current Canvas. Starting the run replaces per-step approval with the visible bounded plan and Stop control.

Agent in a Canvas section

Section Agent stays deliberately small but is a real agent pass: Run Agent authorizes one bounded pass through the same orchestrator that powers the Whole-Canvas loop, focused on the active section and aware of the unsaved draft. Brain plans, improves the draft, shows the visible plan steps in the chat, applies that one update to the editor draft, and stops. It does not start the multi-gap Whole-Canvas loop.

Read-only Brain conversations

AI Readiness, Agentic ROI, Overview, Summary, Slides, and Outputs are read-only conversation surfaces. Brain can explain evidence, critique an artifact, identify missing information, and recommend where to navigate. It cannot claim to edit, save, resolve, generate, or mutate the Canvas from those conversations.

For that reason, these surfaces show a Read-only chat indicator rather than the Chat, Guided, and Agent action dial.

Build mode is a guide, not another autonomy level

Build on the main Brain page sequences the 11 authored Canvas sections and points to the next unfilled section. A section can be skipped for the current session, but that skip is not persisted. Filling every section means the authored sequence is filled; it does not mean every claim is validated, every readiness gap is closed, or the Brain is "complete."

The score shown during Build mode is the AI Readiness score. Section count and readiness confidence are related signals, not the same measurement.

Fresh chats, History, and audit

The Brain opens with a fresh visible conversation. Earlier conversations remain available through History and are scoped to Whole Canvas, a Canvas section, AI Readiness, or Agentic ROI. The Brain history page also shows growth-log events so the team can distinguish conversation from changes to the model.

History provides context and accountability, but an old statement is not automatically current evidence. Correct the Canvas or detach stale memory when the underlying operation changes.

Sources and attached memory

The Sources control accepts notes, pasted excerpts, a Canvas snapshot, and supported Google Drive links. Raw material must be reviewed and approved before its summary and facts become attached memory. See Sources and memory for the lifecycle and limits.

Next: Canvas guide.