With Sonnet 4.6, Anthropic quietly shipped something that I think is way more important than most people realize: programmatic tool calling. The core idea is simple—stop forcing agents to emit verbose JSON tool calls, and instead let them do what they’re actually trained to do: write code. In a sandbox, the agent can orchestrate tool invocations, control ordering, and keep all the intermediate tool I/O out of the context window. You only feed the final result back, which means drastically fewer tokens and less context pollution. In this video I break down why this matters in the bigger “context engineering” problem—especially now that MCP-style setups load tons of tool definitions and tool outputs into the prompt. I walk through the timeline of “code mode” adoption (Cloudflare, then Anthropic) and why I expect this to become a standard the same way MCP and agent skills did. Finally, I cover Anthropic’s improved web search + dynamic filtering, which uses programmatic tool calling to post-process search results before they hit the context window. On benchmarks like BrowserComp and DeepSearch QA, they report meaningful accuracy gains with fewer input tokens—though I also explain why token cost doesn’t always go down if the model writes a lot of filtering code (Opus is the obvious example).