Deep Agents v0.5 Released

Deep Agents v0.5 has been released as a new minor version for deepagents and deepagentsjs. The core addition is async subagents, which allow work to be handed off to remote agents and processed in parallel in the background, instead of blocking the main agent as before.

Now a supervisor can kick off a task and immediately continue conversing with the user or move on to other work. Because async subagents are stateful execution units that retain state, it's also possible to send follow-up instructions or adjust the direction of a task mid-run.

The new interface provides the following five tools:

• start_async_task: Start a task on a remote agent
• check_async_task: Check status and retrieve results
• update_async_task: Send additional instructions to a running task
• cancel_async_task: Cancel execution
• list_async_tasks: View the list of tracked tasks

Any server that implements the Agent Protocol can be connected as a target. This includes agents deployed on LangSmith, custom FastAPI-based servers, and even ASGI transport running in the same process, and example servers for Deep Agents have been added for both Python and JS.

The reasoning behind the protocol choice is clear. ACP is a stdio-centric synchronous session model, so it doesn't fit remote async tasks well, and while A2A supports HTTP and async tasks, it wasn't adopted first since rapid iteration is important at this stage. Agent Protocol, on the other hand, has a thread and run model that fits well with how async subagents behave, and it connects naturally to state retention and delivering follow-up updates.

File-handling capability has also been expanded. Previously, only reading images from the virtual filesystem was possible, but starting with this release, a variety of file formats including PDF, audio, and video can be handled through a single read_file. The file type is automatically determined based on the extension, and it's passed to the model as a native content block of the corresponding MIME type.

However, which modalities are available differs by model. Support can be checked programmatically via model profiles, with each LangChain chat model specifying the input types it accepts in its profile.