Implementing programmatic tool calling on Amazon Bedrock

2026-05-19 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cloud Computing & IT Infrastructure, Software Development & Engineering · Depth: Advanced, long

Summary

Programmatic Tool Calling (PTC) is a paradigm shift in how large language models (LLMs) interact with external tools, moving from sequential, one-at-a-time invocations to code-orchestrated execution. This approach involves the LLM generating Python code that runs in a sandboxed environment, handling multiple tool calls, data processing, and filtering. This dramatically reduces latency and token usage, with experimental results showing an 87-92% reduction in token consumption across models like Claude Sonnet 4.6, Qwen3-Coder-480B, and MiniMax M2.1. Accuracy also significantly improves as data processing occurs in Python rather than natural language. The article presents three implementation methods on Amazon Bedrock: a self-hosted Docker sandbox on ECS for maximum control, a managed solution using Amazon Bedrock AgentCore Code Interpreter, and an Anthropic SDK-compatible path via a proxy.

Key takeaway

For AI Engineers or ML Architects designing LLM-powered agents with complex multi-step workflows, adopting Programmatic Tool Calling (PTC) on Amazon Bedrock is crucial. You can achieve 87-92% token cost reduction and significantly improve accuracy by offloading data processing to a sandboxed Python environment. Consider the self-hosted ECS option for full control over custom packages and security, or the managed AgentCore Code Interpreter for reduced operational overhead, to optimize your agent's performance and cost efficiency.

Key insights

PTC enables LLMs to orchestrate multi-step tool interactions via generated code in a sandbox, drastically cutting tokens and boosting accuracy.

Principles

• Code-orchestrated tool execution reduces LLM context window burden.
• Sandboxed code execution enhances data privacy and processing accuracy.
• Model-agnostic PTC improves performance across diverse LLMs.

Method

LLM generates Python code for tool orchestration; code executes in an isolated sandbox; orchestrator intercepts tool calls, executes them, and injects results; only final output returns to LLM.

In practice

• Implement PTC on Amazon Bedrock via self-hosted Docker/ECS.
• Use Amazon Bedrock AgentCore Code Interpreter for managed PTC.
• Integrate Anthropic SDK with Bedrock via a translation proxy.

Topics

• Programmatic Tool Calling
• Large Language Models
• Amazon Bedrock
• Code Interpreter
• Sandbox Security
• Token Optimization
• AI Agents

Code references

• aws-samples/sample-ai-possibilities
• aws-samples/sample-bedrock-api-proxy

Best for: AI Engineer, Machine Learning Engineer, AI Architect

Related on AIssential

• Build custom code-based evaluators in Amazon Bedrock AgentCore — Custom code-based evaluators in Amazon Bedrock AgentCore enable deterministic, domain-specific quality checks for agentic applications.
• Easy Agentic Tool Calling with Gemma 4 — True LLM agency requires interacting with the local system through carefully sandboxed tools, enabling observation and deterministic computation beyond remote API calls.
• Break the context window barrier with Amazon Bedrock AgentCore — RLM treats context as an external environment, enabling LLMs to programmatically interact with and analyze arbitrarily long documents.
• Anthropic killed Tool calling — Anthropic's tool calling updates enhance agent efficiency and accuracy for complex tasks by optimizing token use and improving tool interaction.
• Semantic search across MCP tools with Amazon Bedrock AgentCore Gateway — Amazon Bedrock AgentCore Gateway enables scalable LLM agents through built-in semantic search for tool selection.
• Tool Calling Is Not an API Call: What Engineers Keep Getting Wrong — Tool calling in LLM agents demands robust engineering for reliability, unlike simple API integrations.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.