Anthropic Just Changed How Agents Call Tools. I Stole It for My Qwen3.5 Agent
👉 Get ALL of our systems & join hundreds of AI builders in our community
https://www.theaiautomators.com/?utm_...
🔗 PRDs: GitHub Repo: https://github.com/theaiautomators/cl...
What if your AI agent could cut its token usage dramatically by searching for tools on demand and writing code to orchestrate dozens of tool calls, instead of letting the LLM handle each one individually?
In this video, we implement two advanced tool calling patterns from Anthropic's engineering post into our custom Python and React AI agent: the tool search tool and programmatic tool calling via code execution sandboxes.
These aren't Claude-exclusive features. They're universal agent design patterns you can use with any model or framework.
We test both approaches side by side, comparing traditional tool calling against programmatic execution, and we run the same task on Claude Haiku and the brand new Qwen 3.5 27B running locally on an RTX 5090.
🔗 Links:
GitHub Repo: https://github.com/theaiautomators/cl...
Anthropic Advanced Tool Use Post: https://www.anthropic.com/engineering...
LLM Sandbox: https://github.com/vndee/llm-sandbox
Cloudflare Code Mode : https://blog.cloudflare.com/code-mode/
Episode 4 (Skills & Sandboxes): • Are Agent Skills the New RAG?
Full codebase available to AI Automators community members
📌 What's covered:
The problem: tool definitions eating context before you even send a message
Tool search: deferring tool loading so the agent discovers what it needs on demand
Programmatic tool calling: letting the LLM write and execute code to orchestrate tool calls in a loop
The tool bridge architecture: how the sandbox calls tools securely without direct internet access or API credentials
Side-by-side comparison of traditional vs programmatic approaches on a budget compliance task (20 team members, 3 tools)
Real results: traditional approach missed answers at 76K tokens; programmatic got them all correct
Running Qwen 3.5 27B locally and comparing it against Claude Haiku
Tool use examples: multi-shot prompting for complex parameter handling (72% to 90% accuracy improvement)
When to use each pattern: context bloat, intermediate result pollution, or parameter accuracy issues
🔍 Tech stack:
Python backend / React frontend
LLM Sandbox (Docker-based isolated code execution)
Langfuse (observability and token tracking)
Qwen 3.5 27B via Ollama (local, RTX 5090) vs Claude Haiku via OpenRouter
Tool bridge pattern for secure sandbox-to-API communication
gVisor (recommended for production sandbox hardening)
Key takeaway: Traditional tool calling breaks down as you scale to dozens of tools and hundreds of data points.
Tool search keeps your context lean, and programmatic tool calling lets the agent write efficient loops instead of burning tokens on repetitive LLM round trips.
These are model-agnostic patterns you can build into any agent system today.
📌 This is Episode 5 of our AI Builder series where we're building a full AI agent web app from scratch using Claude Code.
⏱️ Timestamps:
00:00 Intro & Tool Search
03:56 Programmatic Tool Calling
10:19 Architecture Overview
15:47 Tool Use Examples
#AI #ToolCalling #AIAgents #CodeExecution #Docker #Sandboxes #ClaudeCode #AgenticRAG #LLM #MCP #Qwen #Anthropic