Deep Agents Introduces Interpreter Skills for Deterministic Code Execution

Deep Agents now supports interpreter skills, a new extension that bundles a TypeScript module with a skill so the agent can import and execute deterministic code during a conversation. The interpreter runtime runs alongside the harness, allowing the agent to spawn subagents, call tools, and maintain state across turns while still being governed by an allowlist. Skills remain the instruction surface, but the interpreter module provides an API surface that the agent can invoke when the skill is relevant.

The article demonstrates a github‑triage skill that imports a triage function, fetches open GitHub items, spawns subagents for each item, and returns a structured result that the agent can further process. Interpreter skills enable deterministic workflows such as repo triage, reducing the model’s need to remember hundreds of partial steps and mitigating context anxiety. The interpreter can expose helper functions like result.toMarkdown() for rendering, and the harness can expose filesystem, network, and tool access selectively. This new capability makes agent work more reliable, testable, and reusable by moving logic into reviewed code rather than loose instructions.