Agents draft, you decide

GrowthOS is built on one division of labour: agentsAgentAn automated worker inside GrowthOS. Agents research your market, propose opportunities, write drafts, and score the pages you already have. Every agent stops at a proposal: nothing publishes and nothing moves forward until you review it and decide., the automated workers that research, draft and score, do production, and you do judgement. Every stage produces a proposal and stops. Nothing advances on your behalf, and nothing publishes itself.

Some runs do start on their own: Setup AutopilotSetup AutopilotThe chain of agents GrowthX runs before you first sign in. It researches your company and market from public sources and drafts your context. What you meet is the result, not the run, and correcting that result is your first real task in a new workspace., the chain of agents that drafts your context before you first sign in, runs when a workspaceWorkspaceEverything GrowthOS knows and does for one brand. Every figure, page, and agent in your workspace covers that brand only. (everything GrowthOS knows and does for one brand) is created, and your GrowthX strategistStrategistYour GrowthX contact: the person who configures your workspace, runs the parts of the product that sit outside it, and takes the corrections only you can catch. triggers analysis runs. What they produce is still a proposal waiting on you.

That constraint is what makes it safe to let an agent write. It's also the thing most likely to make the product feel slow if you were expecting automation, and the thing that makes it compound if you work with it rather than around it.

Speed bought with assumptions

A new workspace is usable within hours instead of after a research engagement. Autopilot buys that speed by making assumptions from public sources, and the price is that some of those assumptions are wrong in ways only you can see.

The failure mode is specific: every stage produces plausible prose. Nothing in the product distinguishes an agent's guess from a human-confirmed fact. Research can only find what is public, so the gaps are predictable:

• positioning you have changed recently
• the customers you actually convert, as opposed to the ones you target
• anything you say in sales calls but never publish

That is why the first real task in a new workspace is reading Foundation, the 4 base contextContextWhat GrowthOS knows about your company: what you do, the market you sit in, the buyers you write for, and how you sound. Held on 8 screens, and every agent reads context before it does anything. documents (Company Overview, Ecosystem, Product & Features, Ideal Customer), end to end rather than skimming it.

Reading is the job here, because you are the only one who can catch what is wrong. When you find something, correct it yourself: each document opens in an editor, you type over the wrong claim, and you click Done to save it. Last edited changes from Agent to your name. When a document is wrong from top to bottom, ask your strategist to regenerate it instead of editing line by line. Either way, that correction is worth more than any single page edit, because every later draft reads from it.

Ready does not mean checked

A document's status in Context reads Ready when it has text in it, and nothing more. GrowthOS sets that status when the text lands; it does not record that a person has read it.

The Last edited column is the review log: anything still attributed to Agent has never been looked at by a person. Track your own review pass, because the product will not track it for you.

Annotation is the training signal. Editing is not.

This is the single habit that separates a workspace that improves from one that stays generic.

When a draft is wrong, you can silently rewrite it or you can leave an annotationAnnotationA note you leave on a draft saying what is off and why, instead of silently rewriting it. Annotations accumulate as pending feedback and fold into your writing profile when someone clicks **Update profile**. Annotating is what teaches the next draft; editing is not., a note saying what is off and why. Both produce a correct page. Only one of them teaches the next draft.

The first few pieces cost real editorial time either way. Put that time into annotations and the pieces that follow need less of it. Put it into silent rewrites and you're still paying the same price 40 pieces later.

And annotations do not train anything until someone presses the button. They accumulate as pending feedback and sit there. They fold into your voice when a person clicks Update profile in the Writing Profile area, the compiled version of how you sound, which is a deliberate act, not a background process.

This is the most common way the loop quietly fails to close. A workspace can carry 50 careful annotations and still draft exactly as it did on day 1, because nobody ran a compile. Annotating without compiling produces the work of teaching and none of the learning. Run one after every few reviewed articles.

The same logic applies to Competitors, where research answers "who ranks near you" and only you can answer "who shows up in deals". The difference between those two lists is exactly where your calibration earns its keep.

Publishing is always a separate act

Ship It prepares everything: title, slug, meta, and a cover image generated from the styles your strategist configures.

Pushing it live is its own step, taken by you. If your workspace has a CMS connection, GrowthOS stages the article in Sanity, Webflow or WordPress as a draft; you publish it there. Otherwise you take it out as markdown or an export bundle. Either way, you paste the live URL back to mark the briefBriefThe plan for one page: the research, structure, and targeting a draft is written from. (the plan the page was written from) Published.

Where this shows up

• Setup Autopilot, the run your strategist operates before you first sign in, is what makes the assumptions.
• Foundation is where you check and correct them.
• Writing Profile is where annotation compounds into voice.
• Page editor is where the proposal-and-wait pattern is most visible.