One Model Provider Is a Toy Nowadays. Part 4/5

Still the same hook. Claude helped write this. I gave it access to our docs at docs.backboard.io, it read them, and it wrote about what its own underlying API doesn't support. Consistent across all four posts. That's either a good sign for AI honesty or a really clean piece of product marketing. Probably both. You've built the demo. It works. One model, one assistant, answers questions. And then you try to build something real. Real agents don't just answer questions. They take action. They call tools. They run in parallel. They use different models for different tasks. They work across days, not sessions. The Anthropic API is stateless. You already know this. But in this post it matters in a specific way: statelessness is especially painful for tool-calling agents, because tool calls are inherently multi-step processes. The model calls a tool. You run the tool. You send the result back. The model decides what to do next. If any state gets lost along the way, the agent breaks. On the r