DeepSeek's Harness Team Open Sources an Agent That Tends Its Own Bugs
The DeepSeek Harness team has published a design write up on what it calls a Super agent, an always on system that carries several coding tasks at once, finds work without being asked and recovers when a step fails, alongside a research prototype on GitHub under the MIT licence.
The release is explicitly framed as research rather than product. The repository sits in DeepSeek's own GitHub organisation under the subtitle towards a more autonomous agent system, and it had drawn about 4,100 stars and 374 forks by the time of writing. Its readme describes a proactive coding agent built for a single GitHub repository and packaged as a bundle for DeepSeek Harness, and it states that some of the designs in the write up are not implemented in the code that shipped.
But the architecture is the interesting part, because it is a direct answer to a problem every team running coding agents has hit. A single long session accumulates context until it slows down and loses the thread, and a swarm of agents with no supervision duplicates work and reports nothing useful.
The team splits the job in two. A resident Router session takes in every human message across channels and does nothing except dispatch, and the write up says it may not even carry a Bash tool. Isolated Worker sessions do the actual work in their own directories, run concurrently, and report back to the Router rather than talking to users.
Follow up events travel a different path. Each Worker subscribes to the threads it cares about, a pull request or a chat topic, so a review comment or a new message reaches that Worker directly instead of being re judged by the Router every time.
And the whole system is event driven. Chat messages, emails, review comments and continuous integration results arrive as triggers, filtered first by deterministic code and then by a model, which keeps the cheap decisions cheap. The team also separates agent state from execution environments, so the agent can stay online in the cloud while reaching a user's own machines, and runs an external Watchdog session that checks whether Workers are actually making progress rather than merely reporting it.
The behaviour they describe in their own internal user group is the most concrete part. One instance follows a standing rule to reproduce reported bugs silently and speak up only once it has a simple fix ready, and it has sometimes opened a pull request before anyone tagged it.
The numbers are preliminary but specific. On synthetic workloads of about 300 messages carrying roughly 20 work requests, the system finished half the tasks within 12.4 minutes and 95 per cent within 15.7 minutes, ahead of a single session using a subagent tool, at the cost of more cached tokens. Removing the direct routing of follow ups to Workers raised Router decisions by about 30 per cent, and batching under load cut inference time per Router decision from 6.4 seconds to 1.6 seconds.
But the team lists its own open problems, and they are the ones that decide whether this style of agent survives contact with production. Scaling the Router is unresolved, long term memory is unresolved, communication between agents is unresolved, and so is the risk of prompt injection when a cloud agent can reach credentials sitting on a local device.
That last risk deserves attention because the design deliberately blurs a boundary that security teams treat as load bearing. An agent whose state lives in the cloud and whose hands reach into a user's laptop, listening to chat and email, is a large target with useful permissions.
The DeepSeek Harness team published a Super agent design write up with a prototype under the MIT licence
A resident Router dispatches work while isolated Workers execute it in their own sessions and directories
Workers subscribe to pull requests and chat topics so follow up events skip the Router
Half of test tasks finished within 12.4 minutes, 95 per cent within 15.7 minutes
The team flags Router scaling, long term memory and prompt injection as open problems
Source: Pandaily
