DeepSeek Harness releases research prototype of autonomous coding super-agent

A proactive coding assistant

The DeepSeek Harness team recently published a design write-up and a research prototype for what they call a Super-agent. Unlike standard AI assistants that wait for a specific prompt to act, this system is designed to be 'always-on.' It proactively looks for work and attempts to move several tasks forward at the same time.

The prototype, released under the MIT license, is specifically built for coding environments. It can listen to activity on GitHub or the Feishu platform to identify when a user needs help. Once a task is identified, the agent opens a work item, handles the logic in its own dedicated session, and then reports back to the user.

For example, if a user reports a bug in a chat, the agent can work on the fix independently. It then communicates the results back to the original conversation, often by opening a pull request. As reported by pandaily.com, the team emphasizes that this is a research project and not yet a finished commercial product.

The Router and Worker design

The technical core of the Super-agent relies on a two-layer architecture to manage complexity. The first layer is a resident Router session. This component acts as the primary interface for all human interactions across different channels. Its sole job is to take in information and dispatch tasks to the appropriate places rather than doing the heavy lifting itself.

The second layer consists of isolated Worker sessions. These workers are the ones that actually perform the tasks using full tools. Because they are isolated, they can run concurrently without interfering with one another. They subscribe to specific threads, such as a particular pull request or a chat topic, so they can receive follow-up events directly.

This separation allows the system to remain organized even as the workload grows. By keeping the Router focused on coordination and the Workers focused on execution, the team can manage multiple threads of work. The system is event-driven, meaning triggers like emails, review comments, or CI results automatically kick off the necessary actions.

Autonomy and progress tracking

To ensure the agent stays on track, the team separated the agent's state from its execution environment. This allows the agent to remain online in the cloud while it reaches out to a user's local machine to perform actions. It also includes an external Watchdog session that monitors the Workers to ensure they are making actual progress on their assigned goals.

The team tested this approach on synthetic workloads involving about 300 messages and 20 work requests. In these tests, the Super-agent finished half of the tasks in 12.4 minutes and 95% of the tasks within 15.7 minutes. This was faster than a single session using a subagent tool, though it did require more cached tokens to achieve the result.

The results also showed that batching helped lower the inference time per Router decision. Specifically, batching under load reduced the time from 6.4 seconds to 1.6 seconds. These findings serve as a preliminary reference for how the team intends to scale the system for more complex, real-world coding environments.

Real-world behavior and risks

In internal testing, the team observed the agent following specific rules to improve its utility. One instance was set up to silently reproduce reported bugs and only speak up once it had a simple fix ready. In some cases, it would even open a pull request before a human had even tagged it for review, demonstrating a high level of autonomy.

However, the team identifies several significant hurdles that remain before this can be widely deployed. They are currently looking at how to scale the Router session to handle more traffic. They also need to solve problems regarding long-term memory and how different agents can communicate with one another without creating errors.

Security is another major concern for the developers. They noted the risk of prompt injection, which is when a user provides input that tricks the AI into performing unintended actions. This is particularly risky for a cloud-based agent that has the ability to reach credentials on a user's local device.