Anthropic Just Changed How Agents Call Tools. I Stole It for My Qwen3.5 Agent

Demonstrates Anthropic-style advanced tool-calling patterns—tool search and programmatic tool calling—in a custom Python/React agent. Shows reducing context bloat by deferring tool schemas, executing code in an isolated sandbox, and routing secure tool calls through a tool bridge. Compares cloud and local models (Claude, Quinn 3.5) and measures token and call efficiency.

– Tool search: defer loading tools and discover by keyword to cut context.
– Programmatic tool calling: LLM generates and iterates on code executed in a sandbox to avoid intermediate token bloat.
– Secure tool bridge: sandboxed code routes external requests via authenticated API stubs, keeping secrets out of the sandbox.
– Tool-use examples and efficient MCP design improve parameter accuracy and scalability.

Quotes:

You don’t load everything up front — defer loading and let the agent search for tools.

Programmatic tool calling: the LLM generates and iterates on code inside an isolated sandbox.

The sandbox has no internet access — every external call goes through a secure tool bridge.