Claude Code adds experimental Code Mode

Replacing sequential JSON tool calling with code execution is the most important architectural shift for coding agents this year. Letting the model write small scripts to orchestrate its own tools eliminates the latency tax and context bloat of multi-turn tool loops.

• –Single-step batching: Claude can chain commands, loop over directories, and execute parallel tool calls in one script instead of dozens of iterative turns
• –Context efficiency: Local filtering and transformations keep massive terminal outputs out of the prompt window, preventing context degradation
• –Opt-in experimental rollout: Enabled via CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 in shell or user settings, supported on Opus 5.5 and Sonnet 5.5
• –Emerging agent paradigm: Mirrors a broader industry shift toward executable sandboxes as the preferred interface over step-by-step function calling