Drafts are cheap. Review is the bottleneck.
My agent produced 43 artifacts in one afternoon. It took me until Friday to read them. That's the design problem we're actually solving.
Tuesday afternoon I pointed Claude Code at a week's worth of customer tickets and asked it to draft response emails, log the themes in our tracking sheet, and create follow-up tasks for anything that felt systemic.
43 minutes later it was done. It had produced:
- 17 email drafts
- 14 rows added to the sheet
- 9 tasks
- 2 notes with the synthesis
- 1 slide I didn't ask for but was honestly useful
I didn't finish reviewing any of it until Friday.
This is the actual shape of the problem
Three years ago the bottleneck was getting a model to produce something halfway usable. Not anymore. The models are astonishingly good. A well-prompted Claude will out-draft me on every one of those 43 artifacts on first try.
The bottleneck moved. It is now in the other direction. When a single prompt can produce a week's worth of reviewable material in 43 minutes, the human on the other side becomes the limiter. Not on writing. On reading.
This is the thing I wish more AI-native products were designed around. Every product demo shows you the prompt-to-output flow and calls it done. In real use, that's the cheap half. The expensive half starts the moment the agent finishes.
What review actually needs
If you treat review as the first-class problem, the design constraints end up pretty different from "build a better chat window."
Everything lands in the same place. If one agent output goes to Gmail, another to Sheets, another to Notion, another to Slack, you're doing the triage across six apps. Unreviewable at volume. Spatio puts every artifact an agent produces into one workspace. Mail, sheets, notes, tasks, files. Same surface.
Diffs, not blank pages. A human reviewing a sheet with 14 new rows should see a diff: what changed, who changed it, when. Not a refreshed page they have to scan for differences. We store every edit as an addressable change, and Desktop surfaces them as diffs.
Accept and reject, not read and rewrite. The default interaction on an agent draft should be "looks good, send." The fallback should be a quick edit, not "open this in Word and retype." Drafts ship with an explicit approval step for anything that leaves the workspace. Send mail. Post in chat. Publish a file. Somebody clicks a button.
Reversibility. Every agent action is recorded and reversible from Desktop for 30 days. Because the price of reviewing 43 things quickly is that you'll approve something you shouldn't, and you need to be able to take it back without a war room.
Nothing auto-sends. This one we argued about. You can imagine a world where high-confidence drafts auto-send and only borderline ones wait for review. We're not there. For now, anything going to a customer, teammate, or external service waits for a human on the team to click "send."
Why Spatio Desktop exists
The short version: Desktop is a Gmail-shaped triage surface for agent output.
When your agent uses SpatioMCP to create_note, update_row, create_draft_email, the artifact appears in Desktop. Your team sees it in a feed. You open it, you review it, you edit or approve or delete. Done. Back to the feed.
It looks boring. It should. The interesting thing isn't the UI chrome. It's that when three agents (yours, your cofounder's, your designer's) all produce outputs in the same afternoon, you have exactly one place to go see all of it.
What this is not
It is not a chat interface. Your agent lives in Claude Code, Cursor, Codex, wherever you already prompt it. Desktop is downstream of that. It is what happens after the prompt.
It is also not a project management tool. You're not going to assign, estimate, or sprint-plan in Desktop. It is a review layer. If we ever start building sprint tooling inside Desktop, email me and tell me I lost the plot.
What to try
Drop SpatioMCP into your Claude Code config. Give it a real task. "Clean up my inbox." "Update the forecast with this quarter's numbers." "Draft responses to the top 10 tickets." Let it run.
Then open Desktop and watch the artifacts land. That's the moment the whole product clicks. The prompt was easy. Now you actually get to see what's expensive.
Matt