"But desktop agents aren't good for coding"
So... have it use a coding agent.
My @bytebot_ai opens up the terminal, and logs into Claude Code via the browser, and then gets to work.
Running Claude Code within Bytebot unlocks all sorts of possibilities.
Made a demo of Pulp, a framework with a CLI and Claude plugin to translate Figma and other AI-generated designs into native apps and audio plugins. Output targets AU, VST3, CLAP, and standalone apps, runs on Skia + Dawn/WebGPU, and no browser or webview. https://t.co/3gb60K7H0Y
For almost everyone, computers were too complicated for far too long. Smartphones fixed that for people by narrowing what they could do. For the last ~5 months AI has done something similar for people who already knew their way around computers by making the hard parts disappear.
"But desktop agents aren't good for coding"
So... have it use a coding agent.
My @bytebot_ai opens up the terminal, and logs into Claude Code via the browser, and then gets to work.
Running Claude Code within Bytebot unlocks all sorts of possibilities.
High-quality screens in seconds.
Use it standalone.
Export to FlutterFlow.
Whatever works for you.
Oh, and it's free.
https://t.co/cFu1oa6H5K
Let us know what you think!
1/ At Standard Fleet: our iOS app for fleet management represents YEARS of development.
A new engineer ported it to Android in 7 hours.
As CEO, I don't code much anymore—mostly customers and product. This broke my brain.
So I spent the weekend vibe-coding. Here's what I learned:
we are in the early C era of prompting coding agents. Manually manage context (memory) and often drop down into the code (ASM) to debug problems on specific programming languages (Architectures)
What will the python look like? Shared and guaranteed architecture primitives? Building block prompts for high quality test harnesses etc? Locked frameworks and languages? What is a garbage collector? What else?
Cheers @hellovai@GeoffreyHuntley who have both talked about this a lot
hm this “Tesla Master Plan Part IV” doesn’t feel written by Elon at all. Part 1 was so clear, simple, elegant. The quantity of em dashes and genericness feels like it was written by AI.
I built an AI agent that lets me speak to car engines.
Dashboard lights waste time and money, so I built a tool that uses AI to diagnose issues.
I’ve tested it on 40 cars and it’s hitting 95% accuracy. making workflows faster, smarter, and more efficient.
all model companies were pretraining on the ~same internet. of course, grok has access to twitter dataset and gemini can pretrain on youtube & so on. but it's mostly the same internet
otoh, rl envs will be w.e the lab chooses to prioritize. so you should expect more speciation