LangChain introduces 'interpreter skills' for Deep Agents — an extension to the existing skills system that bundles a TypeScript module alongside skill instructions. Instead of relying solely on prompt instructions for agent behavior, interpreter skills let the deterministic parts of a workflow live in reviewed, testable code. The agent still decides when to invoke the skill and what inputs to pass, but the actual procedure runs as code inside an embedded TypeScript runtime. This enables spawning subagents programmatically, managing task graphs, and handling partial failures within a single workflow. A GitHub repo triage example demonstrates how 300+ items can be processed via subagents without the model tracking every intermediate state. The approach improves evaluability — teams can ask concrete questions like 'did the agent call the expected function?' rather than fuzzy ones about whether instructions were followed.
