Implementing programmatic tool calling on Amazon Bedrock

This article explains programmatic tool calling (PTC), a paradigm that enables LLMs to generate Python code that orchestrates multiple tool calls within a sandboxed environment, dramatically reducing latency and token consumption compared to traditional sequential tool calling.

• PTC generates code once; sandbox executes tools; only final result returns to model context
• Reduces token consumption by 87–92% by eliminating intermediate results from context window
• Improves accuracy by using Python for data processing instead of natural language reasoning
• Three implementation approaches: self-hosted Docker on ECS, managed AgentCore Code Interpreter, Anthropic SDK proxy
• Self-hosted solution uses IPC protocol between orchestrator and Docker sandbox for tool execution
• Sandbox enforces strict security: no network, read-only filesystem, non-root user, resource limits
• AgentCore pre-loads tools into sandbox; model generates code calling pre-loaded functions directly
• Tested across eight models (Claude, Qwen, DeepSeek, MiniMax, Kimi, GLM); all achieved correct results in PTC mode
• Cost savings example: 90% reduction ($14,040/month savings on 1,000 daily executions)

PTC is model-agnostic and privately deployed within AWS accounts, offering significant efficiency gains for multi-tool workflows involving data processing, calculations, and orchestration.