LLMs and Code Mode

06-02-2026

LLMs are better at writing code then calling tools:

• https://machinelearning.apple.com/research/codeact
• https://blog.cloudflare.com/code-mode/
• https://huggingface.co/learn/agents-course/en/unit2/smolagents/code_agents
• https://platform.claude.com/docs/en/agents-and-tools/tool-use/programmatic-tool-calling

MCPs are still useful:

• use MCP for standardisation and tool discovery
• use agents to convert MCP tools into APIs; then search tools, generate code, and run it
• https://www.anthropic.com/engineering/code-execution-with-mcp
• https://github.com/pydantic/monty
• https://github.com/universal-tool-calling-protocol/code-mode - a good starting point