DeepSeek’s Super-Agent diagram: The router is cool, but the comments first ask how context is divided

DeepSeek Harness published an article with a diagram explaining how Super-Agent runs. The diagram first contrasts this with the traditional approach: one session per task, switching back and forth among chat, GitHub, and email. Super-Agent changes this to one agent running multiple tasks across platforms.

The architecture is not complicated. A persistent router receives events from chat, GitHub, and email, filters them according to rules, and then dispatches them to the corresponding workers. The diagram gives the example of PR comments: when someone comments on a PR, the router sends the comment to the worker responsible for that PR. If a problem or requirement is mentioned in a chat thread, Super-Agent can also respond automatically. Context management is the key point: most context is stored in an external database and loaded only when needed, so the prompt does not keep getting longer and longer.

A user translated this diagram into Chinese, making the information more intuitive.

Another user posted a TL;DR diagram, essentially compressing routing, workers, and event-driven execution into a single image.

The comments did not stop at “How strong is the model?” One user said router plus isolated workers is like how a good studio operates. Another user focused on per-worker context and argued that one massive shared context is the common breaking point for agents. Another asked more specifically: when multiple workers share one user chat, does each worker read the entire message history, or only the part forwarded by the router? This question goes straight to implementation details.

Others were concerned about how the main agent checks the work of sub-agents and then uses the results. This point is closer to real-world deployment than “it can run in parallel.” Parallelism is not hard to draw; what is hard is who is responsible for the results, at which layer checks happen, and how to recover after failure. Another user judged that as more sub-agents are added, routing and shared context will become bottlenecks first, while a single model will not be the most constrained part.

After the diagram spread, there were also different opinions. Someone asked whether the font could be made more normal, and someone else replied directly: can you not turn a good article into AI slop? On the other hand, a user cut the same architecture into a one-minute video, saving people the step of reading a long article.

The original article is at DeepSeek Harness. The interesting part of Super-Agent may lie in orchestration and context splitting. The most copy-worthy part of the diagram is that the router hands off work, workers each do their jobs with limited context, and the main agent still has to quality-check the deliverables. Model capability is of course important, but the questions in the comments are more honest: after multiple workers, routing, context isolation, and inspection processes will expose weaknesses first.