[Feature]: Programmatic Tool Calling (PTC) for code based tool execution

[therealtimex]

🚀 The feature, motivation and pitch

When using tool calling with LLMs, each tool invocation requires a separate model turn, leading to:

• High latency for multi-tool workflows
• Increased token consumption from repeated context
• Limited ability to perform complex logic (filtering, aggregation, conditionals) across tool results

xref:

• Anthropic's https://www.anthropic.com/engineering/advanced-tool-use
• Anthropic's https://www.anthropic.com/engineering/code-execution-with-mcp
• MCP Server with Code Execution https://github.com/harche/ProDisco

Proposed Feature: Programmatic Tool Calling (PTC) allows models to write Python code that orchestrates multiple tool calls within a secure sandbox. Instead of returning individual tool calls, the model generates code that:

• Calls multiple tools programmatically
• Processes results with Python logic (loops, filters, aggregations)
• Returns only the final result to the model context

Alternatives

Multiple sequential tool calls, current approach, works but incurs high latency and token cost for complex workflows
(copied from vllm-project/vllm#29573)