Cursor Sand: The New AI Agent Set to change Office Automation

The three-way race for the office agent just got its third runner, and the most interesting one might not be the two already shipping.

Cursor — the AI code editor used across two-thirds of the Fortune 500 — is quietly building a general-purpose agent called Sand to take on Anthropic's Claude Cowork and OpenAI's ChatGPT Work. The Cursor Sand vs Claude Cowork framing misses the only thing that matters: what the agent does after it drafts the email.

That is the real story in the report from The Information via The Next Web on 13 July. Sand is the first of the three office agents with a serious claim on the deployment half of the loop, not just the drafting half.

What Sand actually is

Per The Information, Sand is a general-purpose AI agent inside Cursor. It is designed to answer emails and texts, wrangle spreadsheets, organise documents, and reach into engineering work. That last bullet is the interesting one — and so is the company building it.

Cursor is not a new lab. Its code editor, forked from Microsoft's Visual Studio Code, reached roughly $4bn in annualised revenue by early June, roughly double where it sat in February. Two-thirds of the Fortune 500 are using it. Cursor is a code shop first, and that history shapes what Sand can plug into.

Cursor began testing Sand internally in late June, on computing power it started leasing from SpaceXAI back in April. A public launch is not promised. That is the whole status report on the day the news broke.

The deployment moat — why MCP is the actual story

Both Anthropic and OpenAI have shipped their office agents this month. Anthropic pushed Claude Cowork to mobile and web on 7 July. OpenAI launched ChatGPT Work, powered by its GPT-5.6 model, two days later on 9 July. All three agents can summarise a pile of files and hand you back a draft.

Cursor's differentiator is older than Sand. Cursor's agents already plug into Vercel, GitHub, Slack, and similar developer tools through the Model Context Protocol — the open standard Anthropic first published in 2024. Cursor has spent years building on that fabric. Its existing code editor already rides it.

The practical consequence is a different kind of session. A Cowork session can gather your files, summarise them, and draft a new one. A Sand session can pick up that draft and push it — to a Vercel preview, to a GitHub PR, to a Slack channel. The gap between "the agent handed me a markdown file" and "the agent handed me something live" is the entire office-automation story, and only one of the three agents is even pointed at it.

For a freelancer who wants a landing page live rather than merely written, that gap is the product. For an ops team that wants the agent to file the ticket, not just describe it, the gap is the product.

The three-way race, properly compared

The two established entries have a real head start on reach. Cowork and ChatGPT Work are already shipping; Sand is not. What Cowork and ChatGPT Work do not have is Cursor's integration surface, and that gap is not the kind that closes overnight.

The SpaceX question changes the calculus

Two facts about ownership shift the shape of this race.

First: SpaceX is buying Cursor for $60bn. The Sand roadmap now lives inside a Musk-controlled orbit. Whether Sand ever ships outside the company, or whether it gets folded into a SpaceXAI stack, "may come down to Elon Musk," per the report.

Second: the compute Sand is being trained and run on is leased from SpaceXAI. Cursor started that lease in April, before the acquisition was public. The coupling is tight, and it predates the deal.

This is the part of the Cursor Sand vs Claude Cowork story nobody has data on yet. Anthropic and OpenAI ship to the public and take the public's money. Sand's path looks more like an in-house tool that may or may not escape. For a Fortune 500 buyer trying to plan a 2027 stack on top of one of these agents, that uncertainty is the actual deliverable.

There is a competitive angle worth flagging too: Anthropic publishes the protocol Cursor's agents lean on. What that relationship looks like when Cursor's largest integration targets sit inside a Musk company is an open question. The article does not pin it down, and neither does anyone else yet.

How to actually try this stuff today

Sand is not on a public waitlist yet — internal testing only as of late June, with no published beta. So the realistic "how to use this today" answer is the proxy:

• Want a draft-and-deploy agent right now? Claude Cowork is the closest shipped thing. Anthropic shipped Cowork to mobile and web on 7 July. Sign in with an Anthropic account and it runs.
• Want the strongest raw-model story for office tasks? ChatGPT Work went live on 9 July, on GPT-5.6. It is the freshly launched option and the one to benchmark against.
• Want the integration surface — Vercel, GitHub, Slack — that any of these will eventually need? That is Cursor's existing code editor. It is the substrate Sand is being built on top of, and it is already broadly available.
• Want Sand specifically? You are watching from outside. The product has not shipped, the public-launch question is open, and the parent company just changed hands.

That is the honest state of the market on the day the news broke. The thing to watch is whether Cursor opens any external testing before the SpaceX close — a public Sand beta would force Cowork and ChatGPT Work to answer the deployment story instead of just the drafting one.

What this enables — and what it doesn't

The interesting near-term scenarios are not all in the inbox. Three the Sand story actually makes concrete:

• A freelance landing page in one prompt. Sand inherits Cursor's Vercel and GitHub wiring. The "ship a marketing page this morning" loop is closer to reality than "draft a marketing page this morning." Cowork can do the second one today.
• An ops agent that closes tickets, not just describes them. The MCP path from Slack to GitHub to a deployment is the path Cursor already exercises. Sand riding that same path is a deployment agent, not a chat agent.
• An in-house Cursor at SpaceXAI. The compute lease predates the acquisition. The most likely first customer is inside the parent company. That is not a consumer launch.

The near-term limits are also real. Sand is not on a public waitlist. Anthropic owns the protocol Cursor leans on. OpenAI just shipped a faster underlying model. None of those gaps close by reading the leak.

What stays when the agents churn

A year from now, the headlines will likely be about Claude Cowork v3, ChatGPT Work 2, and whatever Cursor ships under SpaceX. The model behind each of them will probably have changed at least twice in that window. The dashboards, the email composers, the spreadsheet panes — those are sticky. The agents target them; they do not build them.

That is the part of any build worth investing in now. The agent that drafts your invoice this quarter will be different from the one that drafts it next quarter. The component they both render into — the invoice screen, the email composer, the table view, the same form working on web, iOS, and Android, one API — is the durable layer. Whatever wins the office-agent race will sit on top of it.

Build the UI layer to survive the agent. The agent will keep churning. The components underneath should not have to.