Trial
Code Execution ("Code Mode") is the pattern of letting an agent write and run code as its primary way of working, instead of emitting many individual tool calls or trying to solve a task in prose. Two overlapping uses:
- Code as the way to use tools: instead of one model round per tool call - each intermediate result flowing back through the context - the agent writes code that orchestrates the tools (loops, filtering, parallel calls) and only the final result returns to the model. Anthropic reported large token savings from this; Cloudflare productized it as "Code Mode", Pydantic AI ships it as CodeMode.
- Code as the way to do the work: for data transformation, calculation, file and document processing the agent writes a script rather than reasoning token by token - more reliable and verifiable, and the natural mechanism behind skills that bundle reusable procedures.
Considerations
- Requires a sandbox: generated code must run isolated (see the isolation discussion in Harness Engineering) - untrusted code plus data access is a prompt injection surface.
- Best fit for tasks with many steps, structured data or many tools; overkill for a single lookup.
- Verify outcomes against objective signals (tests, schema checks) rather than trusting that the code did the right thing.