A fun side project turned into a byte-sized competition. I was chatting with a few friends about building the smallest possible coding agent. Jesse Vincent jumped in first and built tiny-agent, a JavaScript agent using the Claude code SDK. It started around 1.5 KB, then optimized itself down to just 986 bytes. 🔗 https://lnkd.in/gEe7Qu78 I built my own version in Python, called minicode, also using the Claude code SDK. I started at 623 bytes and had it optimize itself down to 405 bytes. 🔗 https://lnkd.in/g2mbwpb6 Both agents use just one tool: bash. Still, they manage to figure out what to do, run multiple iterations, and answer useful questions. What's wild is that minicode is only 5 lines of code. It does cheat a bit by relying on the Claude code SDK, but even so, the result is surprisingly capable for something so small. You can see minicode activating tools, reflecting, and shrinking itself step by step here: 👉 https://lnkd.in/gYUWVeuY It's pretty amazing how much you can do with so little code when the model does the heavy lifting. Can you get your own minified agents doing cool things? And how should we evaluate how useful these tiny tools really are?
Hi Eran, amazing to see how a fun side project evolved into such impressive results! Your work with minicode is truly inspiring. Great job! 🚀
Fun!
Cofounder & CTO at Keyboardio
1mo(I actually used the anthropic SDK that doesn’t have all of Claude Code baked into it, so I had to do my own tool calling. i’m now at 803 bytes :)