Ives van Hoorne

With Biscuit we have a durable execution system where we run the agent harness in the browser, and hand it over to a server when the browser tab closes. Now someone talked to Biscuit in Slack, and since I had the project open, my browser tab ran the agent harness and answered 😂

Almost everything in Biscuit runs in the browser. Code gets compiled using Oxc (wasm), the LSP runs in a web worker, and the agent harness too. This allows for really quick direct manipulation. You can click on any element and edit it without server roundtrips.

Oxc has been fantastic here. We built custom lints based on common mistakes we saw agents make. For example, we have a rule that watches for Tailwind classes and gives a warning to the agent if it did not implement light/dark mode properly.

In Biscuit you can click on any text and edit it, no matter where it comes from: the database, some variable or just static code. This way you can update any content without leaving the preview. Here's a small thread on how we make this possible:

This makes a better UX for the user, but it also significantly makes it easier for agents to debug code. Our agents see exactly where all variables come from, and things like whether an if condition was true or false. That's a huge help with debugging.

In Biscuit you can click on any text and edit it, no matter where it comes from: the database, some variable or just static code. This way you can update any content without leaving the preview. Here's a small thread on how we make this possible:

For non-React cases, we solve it again with the compiler! We detect cases like this, and wrap with __unwrap. __unwrap turns a traced string into a plain string. This is a last resort, because we lose the origin data. We do it for e.g. comparisons, or when we call libraries.

In Biscuit you can click on any text and edit it, no matter where it comes from: the database, some variable or just static code. This way you can update any content without leaving the preview. Here's a small thread on how we make this possible:

The challenge is that React can’t render traced strings out of the box. And comparisons like `str1 === str2` break if either side is traced. So we do two things: - We use a custom version of React (jsxDev) that supports traced strings - We automatically unwrap in our compiler

In Biscuit you can click on any text and edit it, no matter where it comes from: the database, some variable or just static code. This way you can update any content without leaving the preview. Here's a small thread on how we make this possible:

So we want to make it possible to click any UI, and update the text. But text could come from multiple places, like from the database, source code or even from outside of the app (e.g. network request/library). To solve this, we added "traced strings".

I'm excited to finally share what we've been working on! For the last 8 months, we've been building Biscuit: an all-in-one platform for building a business. Product, payments, emails, hosting and operations built in. Today we're opening the waitlist 🎉

Created a new template for CodeSandbox that runs a headless chromium browser. It exposes a WebSocket port which you can use to control it. This is especially useful with CodeSandbox SDK, as it allows you to create snapshottable & clonable browser environments on the fly 😄

It's now possible to access preview URLs for private sandboxes with CodeSandbox SDK! We also expose a new CLI (`csb sandbox preview-tokens`) to manage preview tokens for a sandbox.

Just published a new feature I'm excited about for the CodeSandbox SDK: Sessions API 🎉 This API allows you to create multiple users in a single sandbox, allowing for collaborative editing and sharing sandboxes publicly. + we use the same API at CodeSandbox 😄

This has been one of the top feature requests since we released the SDK and it unlocks some cool use cases: - Collaboration: allow multiple users to connect to the same sandbox/branch and collaboratively edit files - Anonymous previews: create r/o sessions to share files/shells with others

Our first weekly release for CodeSandbox SDK is live! This week we've focused a ton on performance. VM clones are now 2x faster (1.3s) and snapshot resumes are now within a second (0.6s)! Expect some further improvements the next few weeks. Happy holidays & new years! 🎆

We're starting a new chapter with CodeSandbox! Next to the news of moving and marrying, I promise this is the last big news from me this year 😂. This is huge for us. We'll have more resources to build and grow CodeSandbox, and I'm extremely excited about CodeSandbox SDK.

CodeSandbox SDK allows you to programmatically spin up VMs on our infrastructure. You'll get all the things we've built for CodeSandbox, including memory snapshot/restore and VM cloning. I'll share more examples, but you can already play with it here: github.com/codesandbox/....