DeepSeek open-sources Super-agent: one session takes in work, multiple Workers run in parallel, p95 cut by nearly half

DeepSeek has open-sourced a Super-agent prototype, placed in the DeepSeek Harness repository. The data in the main promotional tweet is straightforward: one router session talks with people and dispatches work, while isolated workers process GitHub event tasks in parallel. In its self-test, p95 completion time dropped from 29.5 minutes to 15.7 minutes, compared with using subagents in a single session. Tweet link: https://x.com/kimmonismus/status/2116905923109089285

The structure isn't complicated. The router manages global context, determines priorities, dispatches tasks, and gives unified replies. Workers do their work in their own sessions and directories, and when they're done they answer in the original conversation, for example through pull requests. There is also a Watchdog that independently monitors exceptions, recovery, and progress checks. A user named lifcc drew a flowchart directly: external events come in from GitHub Issues, PRs, CI, and Feishu messages, and the router assigns them to workers that fix bugs, review PRs, and run tests. The image below shows it.

A user named Silicon-based Trainee JAXX put the old and new architectures side by side: the old version was single-session, serially executed, locally dependent, a single terminal window, and a constrained human-machine loop; the new version is multi-agent collaboration, parallel execution, cloud-native, and autonomously guarded. In his reply he located the pain point in human attention bandwidth: what limits the upper bound of Agent productivity is no longer just model intelligence, but also humans' limited attention bandwidth. The image below compares the two architectures.

Another image comes from a user named Higemaru, illustrating a two-layer architecture that separates the router layer and the worker layer, balancing continuity and parallel processing.

The long article from DeepSeek Harness is here, and the quoted tweet is here. But don't rush to treat it as an official product. A user named Aditya Batra pointed out that the top of the README says this is not an official DeepSeek product, nor is it the same code as in the article. In the replies, someone else asked: the README says some features aren't in the prototype, so which part will be added first? This reminder matters. A prototype being able to run a set of ideas doesn't mean it can be directly moved into a team's workflow.

The comment section also wasn't carried away by the 15.7 minutes. Jeff Bruchado said that if two isolated workers touch the same batch of files, 15.7 minutes could turn into an afternoon of merge conflicts. Abhi asked who resolves conflicts when two parallel workers modify the same file. Capslock asked how isolated the workers really are and whether they share state. MartinOL added another point: p95 being nearly halved looks good, but you also need to put the per-task token count next to it; isolated workers each re-read the repo, so faster may also mean a bigger bill. zeeshan asked: when a worker gets stuck on a flaky test, does the router reassign it or wait? The prototype hasn't given answers to these questions yet.

What DeepSeek is trying this time is a multi-Agent organizational approach; whether the coding model is strong is not the focus. One session handles people, multiple workers work in parallel, the router does scheduling, and the watchdog does monitoring. Orchestration may be as important as model intelligence. But the self-tested p95 is only a self-test; conflict handling, state sharing, token cost, and failure retries are all bills that production environments cannot avoid.

Don't rush to move your team's workflow onto it yet. Wait until the prototype has clearly worked out same-file collisions and the cost ledger, then talk about whether Super-agent can save a team attention.