Ship the impossible

Every app is CRUD

Strip the branding off any SaaS product and you find the same four operations. Create a record. Read a record. Update a record. Delete a record. Spreadsheets do it. Salesforce does it. QuickBooks, Notion, Linear, Figma, all of them.

Workflows get layered on top. Approvals, pipelines, notifications. But underneath the workflows, the primitive is still CRUD. That's the shape of software.

Call this a workspace. When I say "workspace" in this essay, I mean any piece of software that does CRUD on data a human cares about.

Automation hit a ceiling

Since the first workspace, engineers have tried to automate parts of it. Automation is a script. A script runs the same steps every time.

This works beautifully until something unpredictable happens. A new type of file. A weird customer. A rule nobody wrote down. The script breaks, and a human has to step in.

Pre-GPT chatbots hit the same wall. Real conversation is unpredictable. Scripted responses can't keep up. Chatbots got a bad reputation, and earned it.

The lesson from both: if your problem has edge cases, a script is the wrong tool.

ChatGPT arrived, and couldn't touch the workspace

Then ChatGPT showed up, and suddenly unpredictability wasn't a problem. You could talk to it about anything. It would figure out what you meant.

There was one catch. ChatGPT couldn't touch the workspace.

The AI reads text and writes text. The workspace lives in databases and files, behind buttons and forms. The two systems couldn't see each other.

So users improvised. They copied context out of the workspace. They pasted it into ChatGPT. They read the response. They typed it back into the workspace by hand.

The human became a USB cable between two systems that couldn't talk to each other.

It worked, because the alternative was nothing. But it was the worst possible integration pattern. Most of the user's time went to plumbing data between tools, not doing the actual work.

AI as a feature doesn't scale

The next move was obvious. Engineers built AI into their workspaces directly. A "summarize with AI" button. A chat box in the corner. A "generate" option in a menu.

This is the pattern nearly every SaaS product uses today, including some of the most polished software being shipped. It's not a story about bad products. It's a story about an architecture that has a ceiling, no matter how well it's executed.

Here's a concrete example. Imagine an accounting app where the AI could help classify documents. Useful feature. Someone adds a "classify with AI" button. Great.

A week later, a user wants the AI to also suggest which clients are at risk of churning. There's no button for that. The AI never sees the request. Someone files a feature ticket. Six months later, if the feature is popular enough, a new button appears.

But the user wanted ten things, not one. For every action the user might want the AI to help with, the engineers have to ship a button. Every button needs a design, a backend, a review cycle. This can't scale.

The root cause is simple. The AI only has access to the capabilities the engineer pre-decided to expose. Everything outside that small surface is invisible. Even if the AI is smart enough to solve the problem, it's caged.

Cursor changed the shape

Programmers lived through every stage of this mess. They invented solutions as the pain got worse.

Stage one. Copy code into ChatGPT. Get a suggestion back. Paste it into the right place in your editor. Hate your life.

Stage two. Cursor put the chat inside the editor. Less tab-switching. Still the USB-cable pattern underneath.

Stage three. Cursor let you reference files from the chat. The AI could pull context itself. Faster. Still "advising" the programmer.

Stage four. Cursor started suggesting where to insert code changes directly in the file. The AI was pointing at the right place, but the programmer still had to click accept.

Stage five. Cursor gave the AI tools to create, read, update, and delete files directly. No more suggesting. The AI was editing the codebase.

Stage five is where the world changed, and most people missed it.

The AI wasn't advising a programmer anymore. The AI was programming. Same codebase, same operations, same files. The human edited through the UI. The AI edited through its own tools. Both were operators. Full CRUD parity.

For the first time, the AI was inside the workspace. Not glued to the edge of it.

AI-Native Workspaces

Nobody has named this pattern yet. People describe it by pointing at examples. "We're building the Cursor for accountants." "We're Claude Code but for lawyers." Clear enough to sell a product, not clear enough to teach.

Here's the definition, in one sentence.

An AI-native workspace is a digital environment where the AI and the human share all CRUD operations on the same data.

That's the whole concept. Everything else follows from it.

A short test. A workspace is AI-native if all four of these are true:

1. The AI can create data in the workspace without waiting for a human to click a button.
2. The AI can update and delete that data through the same path.
3. The human sees what the AI did in real time, through the normal UI, without refreshing or reopening anything.
4. There is a chat where the human can ask the AI to do anything the human could do by hand.

Four yes answers: AI-native. Anything less: not yet.

Why this matters beyond code

Coding was first because programmers build their own tools. Every other role has been waiting.

But CRUD is universal. The shape Cursor and Claude Code found works everywhere. Bookkeeping is CRUD. Compliance is CRUD. Sales pipelines are CRUD. Legal review is CRUD.

As soon as any of those jobs gets a proper AI-native workspace, it will get the same 10x to 50x jump that coding got.

Every SaaS today is trying to glue AI in after the fact. Most will ship a chat sidebar and call it innovation. Every one of those jobs is still waiting for its Cursor moment.

That's why I decided to build Houston.

What Houston is for users

Houston is a desktop app you download for free. You pick an agent and start working with it the way you'd work with a colleague. Chat with it. Check the board. It handles the rest. If your workflow is a bit different, you just tell it. It adapts. Like a senior team member who already knows your best practices.

Agents aren't chatbots. They use real tools, the same way a human teammate would. Read files, send messages, hit APIs, edit documents. When they're done, the result shows up on your board. When they need a decision, they ask.

Builders create new agents and share them online. You install one in a click. The store grows because anyone can contribute to it.

What Houston is for founders

Houston is the platform where you build AI-native products for your customers. You define your vertical. Houston handles everything else.

The workspace, the AI runtime, the chat, the board, the integrations. All of it is already built. You bring the domain expertise: the prompts, the workflows, the industry knowledge that makes your product worth paying for.

We know this works because we did it ourselves. We had a tax product. Six months of traditional development. Hundreds of AI workflows and edge cases hardcoded into code.

One evening of rebuilding on Houston and we had what six months couldn't produce. Not because we wrote more code. Because the AI handled every edge case without anyone shipping a new button.

You ship one product. Each customer's AI adapts it to fit them. No forks per client. No feature flags. No professional services team maintaining custom branches.

When you're ready to deploy, Houston Cloud runs it for your customers. Your brand, your domain. Our infrastructure.

The whole platform is open source. You can inspect every line. You can fork it and run it yourself. Most won't. The point of a platform is that you shouldn't have to.

One product. Infinite versions. Ship the impossible.