Anthropic cut an AI agent's token bill from 150,000 tokens down to 2,000 by telling it to stop calling its own tools. Picture an agent copying a document from Google Drive into Salesforce. With normal tool calls, that whole document travels through the model. Once to read it, again to write it. Every word paid for twice. Anthropic's fix. Don't call the tools. Write code that calls them instead. The code runs off to the side, moves the document directly, And only the final result comes back. That's the 98% and it stacks. Five pending orders buried in 10,000 spreadsheet rows, the code filters first, the model never sees the rest. Sensitive data never even leaves the sandbox. A great agent doesn't hold everything in its head. It writes a script and just reads the answer.
Machine-transcribed from the published render, so acronyms may be cased as ordinary words.