Deep Agents Adds Interpreter Runtime for Code‑Level Agent Composition
Implement interpreters in your agent harness to reduce token usage by up to 35% and keep intermediate state out of model context.
Implement interpreter support in your agent harness, expose needed tools via explicit bridges, and test token savings on your workflows.
Summary
Deep Agents now introduces an interpreter runtime that lets agents write and execute code inside the agent loop, creating a third context surface separate from message history and the filesystem.
The interpreter holds live working values such as arrays, objects, and helper functions, allowing the model to reason over them without adding prompt tokens. By exposing allowlisted tools under a global tools namespace, developers can enable programmatic tool calling as middleware, which has shown up to 35 % fewer tokens on benchmark tasks. The runtime is intentionally minimal—no filesystem, network, shell, or package installation—so only explicitly bridged actions like fetch, read_file, or task can reach the host environment. This design keeps the agent’s action space narrow, improves predictability, and makes debugging easier. The interpreter can be used with any model, not just proprietary ones, and supports multi‑step logic and state reuse across calls. It complements existing sandbox and serial tool‑loop approaches by offering a middle ground that balances control flow and resource usage. Early tests demonstrate that code‑level composition inside the interpreter reduces token consumption while keeping intermediate results out of the model prompt.
Key changes
- Introduces interpreter runtime that agents can write code against within the agent loop
- Interpreter state is a separate context surface distinct from message history and filesystem
- Allows allowlisted tools exposed under a tools namespace for programmatic tool calling
- Shows up to 35% fewer tokens on benchmark tasks compared to serial tool calls
- Runtime defaults to minimal environment (no filesystem, network, shell, or package install)
- Requires explicit bridges for external actions such as fetch, read_file, and task
- Supports multi‑step logic and state reuse across calls without model prompt injection
- Works with any model, not tied to a specific provider