Use programmatic tool calling so large tool results never enter the model's context

Claude Cookbook·medium signal

A key 2026 context lever is keeping bulky tool outputs out of the window entirely: have the agent call tools programmatically (code that processes results) and route heavy sub-tasks to isolated subagents that return only 1,000-2,000 token condensed summaries. The model reasons over distilled state, not raw payloads. This is how long-horizon agents avoid the token exhaustion that kills naive 'dump every result into the chat' designs.