Agents and people should read the same map.

Picture a canvas with eighty notes on it. Months of thinking, arranged in clusters, linked by arrows, edited by a few people and a couple of agents. Now ask an AI agent a simple question about it: what is going on here, and what should I look at first?

Today the agent has two bad options. It can read everything, which is slow and burns its whole context window on notes that do not matter. Or it can guess from a thin summary and a title list, which means it misses what actually changed, what decision is still open, and where an agent left work nobody has checked. Neither option is how a person works. A person glances, orients, and dives into the one corner that matters.

The usual fix is to build the agent an index. Chop the content into chunks, embed them, and let the agent search. It works, and it is also exactly the thing we do not want. That index is a private map. The human cannot see it, cannot correct it, and cannot tell what the agent thinks the canvas means.

We have written before about why Creator Notes is shared working memory for humans and agents. A hidden index quietly breaks that promise. If the agent understands the canvas through a vector store you can never open, then you and the agent are keeping two separate memories of the same work. That is the opposite of alignment. It is two people in a room reading different documents and pretending they agree.

So we are building the index as part of the canvas, not beside it. The canvas is already a human-readable model of the work. We do not replace that model with a machine one. We make the same model readable to agents.

The rule we hold ourselves to is simple. Every concept the agent uses has to show up somewhere a person can see it. The regions the agent reasons about appear in the canvas info panel. The note about who shaped the work appears as a bar you can read. When you rename a region, the agent gets the new name on its next look. There is no backstage.

That gives us one test for every feature on the roadmap. It has to help a person reorient when they come back, help an agent reorient without rereading everything, help a person trust what the agent did, or help an agent avoid misreading the canvas. If it does none of those, it waits.

The field has good ideas about reading large things without reading all of them. Some build a table of contents and let a model walk it. Some cluster a corpus and summarize the clusters. Some show an agent the shapes in view in full and everything else as a summary. The shared instinct is progressive disclosure. Start with a skeleton, then go deeper only where it pays.

We take the ladder and leave the engine, because a canvas is not a document. It is a graph with a layout and a history. So the levels are ours. A one screen outline answers what the canvas is about. A region answers what is in this part of it. A note answers the detail. An agent can settle the first question in one small request, open a single region in one more, and only then read the two or three notes that actually matter.

The first thing we are shipping is not clustering, and not a fancier search box. It is orientation. The outline an agent gets back is built to answer the questions a person asks when they walk up to the board:

• What is this for? The purpose and the themes, in a sentence, not a paragraph.
• What changed? The recent movement, drawn from the canvas history we already keep.
• What needs my attention? Work an agent created that no human has reviewed yet.
• Where do I start? A short list of the notes worth opening first.

That is most of the value, at the least risk, and it leans entirely on signal the product already records. The richer layers, the ones that draw regions on the board, come after the foundation has earned its place.

Here is the part we think nobody else can copy cheaply. Every node and every change in Creator Notes already carries its provenance, the same trail that powers canvas history. So we can tell you, at a glance, how much of a canvas is human work, how much is agent work, and when each was last touched.

But raw authorship lies if you read it naively. A person who asks an agent to draft thirty notes did not write less; they wrote the prompt. An agent draft that a person edited and kept is not really agent work anymore. A bar that shouts “ninety percent agent” would scare you about work you actually own.

So the signal we lead with is not who typed it. It is trust. Who created it, who last changed its meaning, and whether a human has engaged with it at all. The headline number an agent and a person both see is the honest one: how much agent work is still sitting here unreviewed. An agent can read that and slow down before it touches a region nobody has checked. A person can read it and catch the agent reaching too far.

The feature we are quietly most excited about is the one that turns the canvas into memory across sessions. When an agent comes back to a canvas, it should not reread the whole thing. It should pick up where the work left off.

You last worked here yesterday. Since then a teammate added three concerns, one decision changed, and a region was renamed. Start there.

That is a different and better thing than a summary of the canvas. It is the previously on recap before the next episode, and it is the natural twin of the “catch me up” we already built for your own timeline. The same instinct, now pointed at the work an agent and a person are doing together.

The agent’s index should be the human’s view. If the two of you cannot point at the same map, you are not aligned. You are just nearby.

Most systems make the machine’s understanding invisible and ask you to trust it. We think the canvas you can already see is the better place to put it. Make it legible to agents, and keep it visible to people. Everything else is a private index nobody can check.