DeepSeek wants an AI agent that works around the clock on many tasks

DeepSeek's Harness team published a technical paper on October 31, 2026, about a new type of AI agent it calls a Super agent. An agent is an AI system that carries out tasks on its own rather than simply answering questions. The architecture is designed for situations involving many tasks and teamwork.

Criticism of today's agents

According to the team, existing agents can typically be followed across only 3-5 active sessions at a time. The team also notes that agents can get stuck in single sessions, end up with confused context and stop if the local machine goes offline.

The team describes the Super agent as continuous, efficient, proactive and persistent. That is the team's own assessment.

Five main components

According to PANews, the architecture rests on five parts:

- a two-part Router/Worker structure - event-driven design - human-oriented interaction - separation of computation from state and environment - an external Watchdog

The Router handles contact with humans, maintains the overall context and distributes tasks. The Workers carry out the tasks simultaneously, each in its own isolated session. This means one task does not interfere with the others.

According to the team's own description, the agent can communicate over several channels, such as chat and comments on GitHub, and it monitors its own environment. Among the mechanisms the team describes are filtering of triggering events, "Inline & Spill," chained recall, speculative execution and an asynchronous ReAct loop. According to the team, the asynchronous task mechanism has been implemented in DeepSeek Harness but is not yet enabled by default.

Used internally and measured in tests

According to the DeepSeek team, an internal instance is already being used to reproduce issues, respond to code reviews and improve Harness itself.

In synthetic load tests, the Super agent's task time was 12.4 minutes at p50 and 15.7 minutes at p95. p95 means that 95 percent of tasks were faster than that figure. The comparison was against a single-session solution with subagents and a neutral prompt, which had a p95 of 29.5 minutes. The test involved about 300 messages, of which about 20 were work tasks, and was repeated 5 times. The figures are the team's own and were reported via ITHome. It is not clear whether the measurement applies to the internal instance or to a prototype.

A prototype repository exists on GitHub. According to ITHome, it is explicitly not an official DeepSeek product.