Member-only story
Anthropic Shipped Trump Cards and Nobody’s Talking About the Right Thing
Two quiet features just changed how serious AI agents should be built.
Gaurav Shrivastav7 min read·Just now--
This month, Anthropic published two updates that, based on what I’m seeing in developer communities, most people are treating as normal feature announcements.
Programmatic Tool Calling and Web Search Dynamic Filtering are the same core idea approached from two directions. And together, they’re the clearest signal yet of where serious AI agent architecture is actually heading.
Non members can read here for free
Let me explain both and why the underlying logic matters more than the features themselves.
The Tax You’re Paying Without Realizing It
Standard tool use in a Claude agent works like this: Claude calls a tool, the result floods back into context, Claude reads it, decides whether to call another tool, and repeats. Every round-trip re-samples the model. Every intermediate result lives in the context window.
If you’ve built anything with five or more sequential tool calls, you’ve felt this. Token costs compound. Latency compounds. And — this part tends to surprise people — quality can actually go down as context fills with…
