Deep Agents' Interpreter: Code Between Tool Calls and Sandboxes

LangChain has added an interpreter to Deep Agents. The interpreter is a small embedded runtime that lets the agent write and execute code within the work loop, similar to providing a Python or Node REPL.

Existing agents had to choose between sequential tool calls or a full sandbox environment. The interpreter sits at the midpoint between them, allowing the agent to do code-level composition over a limited set of capabilities without inheriting a full environment. By default, there is no filesystem, network, or shell access, and only the capabilities explicitly exposed by the host runtime can be used.

Core capabilities the interpreter provides:

• Using interpreter state as a third context surface: the message history is the context in which the model reasons, the filesystem is for persistent artifacts, and interpreter state stores working values that don't yet need to become model input
• Programmatic Tool Calling (PTC): tool calls happen within code written by the agent rather than in a model-mediated sequence. Implemented as middleware, it's usable with any model, and in early tests reduced token usage for some tasks by up to 35%
• Improved handling of large datasets: the agent can maintain document state in the runtime, programmatically iterate over batches, and invoke subagents only on selected slices

This approach follows a pattern similar to Cloudflare's Code Mode, Anthropic's PTC, and RLM-style workflows, implemented in Deep Agents in a model-agnostic way.