As you add more tools to an agent, it starts calling the wrong ones. Letting the model write code to call them is the fix.

when an agent has a handful of tools, calling them directly works fine. The problems start as the list grows. Every tool schema sits in the context before the agent does any real work, and the model starts reaching for the wrong tool or calling them in a clumsy order. The usual fixes buy a little room, then stop. Shorter tool descriptions, grouping related tools, a stricter system prompt about when to use what. Each one helps a bit. None of them hold once a task needs several tools in sequence. The approach that holds up is to stop exposing every tool to the model directly. Give it a small sandbox where it writes a few lines of code that call the tools as functions, then let it loop, filter, and chain them in code instead of one tool call per step. Two things change. The context stops carrying a wall of schemas, so token use per run drops. And the model makes fewer wrong calls, because writing a short script is something it's already good at. The trade-off is real: you're running model-written code now, so what that sandbox can import and reach is the thing to lock down. If you're running a tool-heavy agent, where's the line between keeping direct tool calls and letting the model write code to call them?