Give Your Agents an Interpreter

Deep Agents introduces interpreters, small embedded runtimes that allow agents to write and execute code within the agent loop, offering a middle ground between one-at-a-time tool calls and full sandboxes. This functionality enables agents to express complex, multi-step workflows, maintain intermediate states outside of the model context, and execute actions more predictably, similar to a Python or Node REPL. Unlike traditional environments, interpreters start with a limited runtime, lacking APIs for filesystem, network, or shell access, and instead rely on explicit bridges for these capabilities. This design ensures a narrower action space and greater predictability, as additional functionalities must be deliberately added by the host. Interpreters facilitate efficient tool calling, manage control flow, and reduce token usage by retaining intermediate outputs within the runtime, rather than routing them back through the model. This setup is particularly beneficial for tasks involving large datasets, programmatic tool calling, and recursive orchestration, allowing agents to manage state more effectively and execute delegated tasks through a controlled API boundary. Deep Agents supports interpreters in both Python and TypeScript, with programmatic tool calling enabled through an allowlist, ensuring that only designated tools can interact with the interpreter.