Farcaster

Had an interesting conversation with colleague who built the earliest visual operator for Google Pixel phone testing. We always thought MCP is overrated for AI, but cannot think of a better way to improve user experience, but a recent paper gives us an idea that could be better than MCP, here is the idea 🧵

1/ Search-R1 is an awesome paper about teaching AI to use search engines through reinforcement learning, we've been thinking: what if we applied this to all tools? Meet Tool-R1, our take on making AI tool use more natural!

2/ Here's the cool part: Search-R1 showed us that AI can learn to search the web naturally through trial and error (reinforcement learning). We thought - why stop at search? Tool-R1 extends this idea to ANY tool - calculators, APIs, you name it!

4/ Think about how you use tools - you don't follow a manual, you just grab what you need when you need it. That's exactly what we should teach AI to do! Tool-R1 makes tool use feel as natural as thinking.

5/ The magic happens in how tools become part of the thinking process. Not as separate steps, but as a fluid part of solving problems. Imagine an AI that grabs a calculator mid-thought, just like you would!

6/ Building on Search-R1's success, I believe Tool-R1 could be a game-changer in how AI interacts with the world. No longer do we need MCP.

Could this be the key to more capable, more natural AI agents?

3/ The big innovation? Instead of hard-coding how AI should use tools (which is what most systems do), Tool-R1 lets AI figure it out through experience - just like humans do! No extra steps, no complex rules - just pure learning from outcomes.