DeepSeek Harness adds plugin for delegating tasks to Claude and Codex

A new way to delegate work

DeepSeek recently published a pre-release version of Harness, its open-source agent runtime. This update introduces an experimental plugin called subagent_oar, which changes how the system handles complex tasks. Instead of trying to do everything itself, the main agent can now hand off self-contained jobs to other specialized tools.

As reported by pandaily.com, the plugin specifically targets Codex CLI and Claude Code installations. These are external tools that exist on the same machine as the Harness host. By using this plugin, the main agent acts as a coordinator that sends instructions to these other systems to perform specific actions.

This differs from previous versions where users had to install these tools as profile bundles. The new method provides a single entry point for dispatching work. When the main agent calls subagent_oar, it chooses which tool to use and submits a task description along with a full prompt and a working directory.

How the hand-off works

The interaction between the main agent and the external tools is designed to be one-shot. This means the main agent sends a request and receives a final answer or an error message back into the main session. There is no live progress stream or ability to ask follow-up questions during the execution of the delegated task.

Because the external tools run as independent entities, they keep their own sign-ins, models, and settings. Users must manually install and log in to each tool to ensure the Harness host can find them. The system can pass specific provider settings like reasoning effort, launch arguments, and environment variables depending on what the specific runtime supports.

The documentation notes that the main conversation is not forwarded to the external tool. This requires the user to be very specific in the prompt about which files matter and what the expected output should be. Only the final result is returned to the main session, meaning the internal reasoning or tool history of the external agent remains private to that specific task.

Workspaces and statistics

The new release includes specific features to help manage these delegated tasks more effectively. For example, official examples show how to enable a writable workspace while turning off permission prompts. This allows the external agent to make changes to files without constantly asking for user approval during the process.

However, there are limitations to how these changes are handled. If a user cancels a task, the system does not undo any file changes that the outside agent has already completed. This means users need to be mindful of the state of their directories when the main agent is delegating work to a subagent.

To help developers monitor these interactions, a new statistics plugin was added. This tool breaks down the usage of context and cache, as well as the time spent on models and tools for every turn and step. It also includes a feature to show the status and differences of a linked pull request if the session is running in its own Git worktree.

Updates and accessibility

Beyond the delegation features, the v0.2.2-alpha.2 release includes several quality-of-life improvements for the Harness platform. It allows users to update third-party plugin packages with a rollback feature in case a hot update fails. It also improves how background-task outputs are displayed within completion notices.

The update also includes guidance for subagent waiting to reduce the amount of repeated polling. This helps the system manage the flow of information more smoothly when waiting for an external tool to finish a long-running job. These technical refinements aim to make the multi-agent experience more stable for the end user.

Finally, the release expands the accessibility of the interface by adding support for several new languages. Users can now interact with the system in Japanese, Spanish, French, German, Russian, and Brazilian Portuguese. These additions make the tool more usable for a global audience of developers and researchers.