Code Mode Agents: Let the LLM write code, not call tools (www.raymondyxu.com)

Anthropic’s “Code Mode” agents shift the agent architecture from repeated LLM→tool→LLM cycles to a single loop where the model writes and executes code that calls your tools. Instead of the LLM prompting a tool, receiving a result, and iterating — which can introduce compounding errors and state drift — Code Mode extracts the tool signatures, generates a runnable script (e.g., agentRunner.py), implements a main() that orchestrates those tools in Python, executes that script, and returns a structured execution result (output, success flag, execution log). The example shows an Architect AI: tools like search_docs, analyze_blueprints, and report_blueprint are exposed, Claude writes helper functions and a main loop to analyze and report blueprints, then the host runs the generated script and returns the aggregated analyses. This matters because it reduces iteration noise and scales far better when an agent must call many functions. LLMs are relatively stronger at generating, verifying and running code than at maintaining long serial tool-call dialogues; Code Mode avoids the error accumulation and context-loss that grows with iteration count and enables essentially unlimited function use in one generated program. Practically, it makes complex orchestration, batching, and offline verification easier. The pattern is already reproducible via open-source tooling (claude_codemode on GitHub) and integrates with existing agent frameworks (pydantic_ai Agent), using models like claude-sonnet-4-5-20250929 to produce and execute the orchestrating code.