New Generation Productivity Suite: AI Artifacts

An AI artifact is a file an AI agent makes for you: an HTML page, a Markdown document, a small app, a live table, an interactive report or a slide deck. You ask in plain language, the agent writes the file, and you get something you can open, click through and send to a colleague. Most AI assistants can now show these files in a panel next to the chat, so you see the result while you talk to the agent.

We think artifacts are the next productivity suite. Word, Excel and PowerPoint have defined office work for more than thirty years. The files people now ask their agents for are built on older and more open formats, and for the first time anyone can make them.

Why are AI artifacts taking off now?

The web has had the right formats for a long time. HTML can lay out anything, run code, draw charts and respond to clicks. Markdown gives you clean documents from plain text. The catch was always the same: you had to be a developer to use them well. A product manager who wanted an interactive pricing calculator wrote a spec and waited for an engineer. Everyone else made do with a spreadsheet and a slide.

AI coding changed that quickly. In February 2025 Andrej Karpathy coined "vibe coding" for building software by describing it and letting the model write the code. By the end of that year it was Collins Dictionary's Word of the Year. Lovable, one of the tools built around the idea, reported $100M in annual revenue eight months after launch, and it's aimed squarely at people who don't write code.

Business users are now the ones asking. Analysts, founders, ops leads and lawyers ask their agent for a dashboard or a one-pager, and they get a working file back in a minute. These formats have been around for decades. Now the people who were locked out of them can use them.

There's a simpler reason too. People communicate visually. A chart you can hover over explains a trend faster than a paragraph does, and a page with sections and a table of contents gets read where a wall of chat text gets skimmed. Once someone gets a proper page back from their agent instead of a long reply, they rarely want to go back.

What is an HTML artifact?

HTML is the language every web page is written in. Tim Berners-Lee described its first 18 tags in a document called "HTML Tags" in 1991, and <p>, <a> and <h1> still work in every browser today. That age is a big part of its appeal. An HTML file opens anywhere, on any device, with nothing to install, and pages written in the 1990s still open today.

HTML gives an agent the fewest limits on what it can hand you. One file can hold styled text, images, charts, tables you can sort, and buttons that do things. Ask for a board update and you can get a page with the numbers, a revenue chart and a short commentary, laid out the way a designer would do it.

What is a Markdown artifact?

Markdown is plain text with a few symbols for formatting: # for a heading, ** for bold, - for a list. John Gruber released it in 2004, with Aaron Swartz as its main tester, because he was tired of writing raw HTML for his blog. It's now the default format of GitHub, of note apps, and of the agents themselves. AI labs write most of the instructions that steer their models as Markdown files. Anil Dash wrote a good history of how Markdown took over the world if you want the full story.

Markdown is the right artifact when the words matter most: meeting notes, a strategy memo, a spec, a research summary. It reads fine as raw text, it renders into a clean document, and both people and agents can edit it without breaking anything.

What is an app artifact?

An app artifact is a small piece of software built for one job. It's still made of web code, but it holds data and lets people act on it. Think of a tracker for 40 investors in a fundraise, a triage screen for a new regulation across every client, or a status page where five workstream owners update their own rows.

Teams used to need an engineer or a no-code tool for these, so most of them never got built. People kept the work in spreadsheets and email threads instead. Now the person who knows the process describes it, and the agent builds an app that fits it. When the process changes next month, they ask again and the app changes with it.

What is a live table?

A live table is a shared list of records that people and agents both work in: a lead list, an inventory count, a log of every customer call. It looks like a spreadsheet when you open it. Underneath, it's a SQL table.

We picked SQL on purpose. The language dates back to the 1970s, and today it holds most of the world's business data, from bank ledgers to online stores. A single Postgres table can hold billions of rows and, with the right index, still answer a query in milliseconds. Spreadsheets give out long before that (Excel stops at 1,048,576 rows), and they get slow well before they hit the cap.

SQL is also one of the languages AI models know best, because fifty years of it is written down in docs, forums and code. An agent that wants every deal that went quiet last month writes one SELECT and gets those rows back. It doesn't download a file, parse it and guess what the columns mean. Adding a record is one INSERT, and changing one is one UPDATE.

In Anchor, every table is a PostgreSQL table. Your agent creates it, adds columns, and reads and writes rows in plain SQL. You open the same table in the browser and watch the changes come in without refreshing. That's what makes it live: you, your team and your agents all work on one copy.

What are interactive docs and slides?

Interactive documents and slides sit between a report and an app. A deck written in HTML can hold a live chart, a filter the audience can play with, or a model where you change an assumption and watch the forecast move. Developers have made slides this way for years with tools like reveal.js (it dates back to 2011). Agents make it available to everyone who presents.

The difference shows up in the meeting. With a static deck, somebody asks "what if churn is 5% instead of 3%?" and the answer is "I'll follow up." With an interactive one, you change the number on screen and everyone sees the result.

Where should AI artifacts live?

Here's the problem we kept running into ourselves. Agents are good at making artifacts, but the artifacts end up in the wrong places. They're stuck in a chat thread, downloaded to a laptop, or hosted by one AI vendor where only that vendor's agent can open them again. Nobody else on the team can find them, and the next agent can't build on them.

We built Anchor as a file system for these files. HTML pages, Markdown docs, apps and live tables sit in files and folders, like on your computer, and every one of them has a link. You make them with the AI you already use, whether that's Claude, ChatGPT or another agent that speaks MCP, so you aren't tied to a single vendor.

That changes two kinds of work.

Agent and human. Your agent writes an artifact into a folder, you open it, and you tell the agent what to change. The next day a different agent can read the same folder and pick up where the first one stopped. The files hold the context, so nobody has to paste it back into a chat.

Human and human. You share the link with your team or with a client, and they open the page itself, charts and buttons included. No screenshots, no attachments. When the data changes, everyone sees the new version. Nobody asks which file is the latest.

We think this is the new way of working. Artifacts are visual, so people understand them faster. They're interactive, so people answer their own questions instead of sending one more email. And they're expressive enough to fit the job, where a slide or a spreadsheet used to be the only option. The hours this saves come from all the small things: fewer status meetings, fewer rebuilt spreadsheets, fewer "can you send me the latest?" messages. That's what productivity looks like for us, and it's why we're building Anchor.