Anthropic API

Following Part 1’s tool-calling loop, Part 2 dives into Agent system prompt architecture: Identity, Instructions, Output, Examples, and Context layers, plus Anthropic Prompt Caching and prompt fingerprinting for regression testing.

A deep dive into the core mechanism of AI agents: the tool-calling loop. Through a ~200-line Rust implementation, we break down how models gain the ability to act via tool use, and why the architecture behind Claude Code and GitHub Copilot is the same fundamental loop.