The warning comes first
Most agent launches bury their limits in a footnote. This one opens with them. The first lines of the dsh-superagent README call the code a "research prototype for reference only, not an official DeepSeek product," note that coding agents did much of the engineering, and warn that it may still contain bugs. The README makes a second distinction that matters just as much: the published code is not identical to DeepSeek's internal implementation or to the system described in its accompanying design article, Exploring Super-agent Design. Some of the designs in that article are not implemented in this prototype. The download is therefore a reference implementation of selected ideas, not the internal system in miniature.
Some coverage has described the broader DeepSeek system as an autonomous coder that watches team conversations, identifies bugs and proposes fixes without being explicitly tagged. That description belongs to the internal system and the accompanying design narrative. The released prototype is narrower: it handles permitted requests and team-defined rules, rather than shipping every proactive capability described in the design article. The distinction matters because the broader design narrative describes capabilities beyond those implemented in the published repository. Readers evaluating the download should judge the code that is actually available, not the more ambitious system described in the accompanying research.
How the released prototype works
The prototype is a bundle for DeepSeek Harness, the lab's agent runtime. It listens to GitHub and, optionally, Feishu. An authorized request becomes a Ticket, a unit of work carrying a Mandate that records why it may proceed. A long-running Router handles incoming events and dispatch; each Ticket runs in its own Worker session and directory, and the result returns to the original conversation, often as a pull request. A web panel lists the items requiring human attention. These are separate execution contexts, not independent OS security boundaries.
The design principle is stated plainly in the README: take as little human attention as possible, and give as much value as possible. The accompanying design narrative argues that the scarce resource in a team running several agents is not model capability but supervision: one person can follow only a limited number of sessions at once. The published prototype addresses that problem through routing, authorization and a panel that surfaces only work requiring human attention. The gating reflects that. Proactive behavior has three settings, off, observe and auto, and the default is observe, in which the agent records what it would have accepted and does nothing. Autonomy, in the shipped code, is opt-in.
The speed figure, read carefully
DeepSeek's reported test used a simulated stream of about 300 messages, including about 20 work requests, with each configuration run five times. The Super-agent finished half its tasks in 12.4 minutes and 95% within 15.7 minutes. A single-session agent with subagents recorded a 95th percentile of 29.5 minutes. These are DeepSeek's reported results, not an independent performance comparison, and the two p95 figures should not be read as a general speed multiplier.
The baseline also needs context. Under a neutral prompt, it reportedly called a subagent zero times in all five runs. The comparison therefore shows how the prototype's routing architecture performed against that particular baseline configuration, not how much faster it is than a well-tuned system that actually parallelizes work. The simulation measures task handling and queue movement, not real-world coding quality.
The same reported experiments include an ablation: removing direct event pass-through increased cache-read tokens from 23.23 million to 31.77 million, approximately 37%. That is a result from DeepSeek's reported test, not an independently reproduced measurement. It illustrates the potential cost of repeatedly passing context through an agent system, but should not be treated as a general token-saving guarantee.
What to do with it
The repository is free under MIT and small enough to inspect: one commit, with numbered design documents covering philosophy, authorization, execution and deployment. For a team that wants to study router-and-worker orchestration, it is a readable reference implementation with its assumptions written down. The security warning deserves particular attention. The deployment documentation says both Router and Worker sessions run with danger-full-access under the operator's OS account. A separate Harness profile does not isolate that access: a worker can execute programs and read files in the operator's home directory. DeepSeek recommends a dedicated OS account containing only the agent's credentials. Treat that as a minimum precaution, and use an appropriately isolated environment rather than pointing the prototype at a machine or repository that matters.
What it is not, yet, is evidence that an unsupervised agent can be left in a production codebase. DeepSeek has published an architecture and a simulation. The claim that would change purchasing decisions, that this system fixes real bugs at a quality a team can merge, has no measurement behind it in anything released so far.
Sources
- [1] DeepSeek Harness team, deepseek-ai/dsh-superagent-prototype on GitHub — README, repository metadata, MIT license and package.json (primary source, read first-hand 2026-11-01; created 2026-10-31, one commit, version 0.1.0)Read source
- [2] DeepSeek repository documentation, 002 Authorization — the accept modes off, observe (default) and auto, and how Mandates record why work may proceed (primary source, read first-hand)Read source
- [3] DeepSeek repository documentation, 006 Deployment — prerequisites, the dedicated Harness profile, and Permissions and isolation: both presets run with danger-full-access as the operator's OS user (primary source, read first-hand; 000 Philosophy, 001 Domain Model and 007 Architecture also read first-hand)Read source
- [4] MarkTechPost, "DeepSeek Open-Sources dsh-superagent, a Router-Worker Coding Agent", November 1, 2026 (secondary reporting of the design post's performance figures, including ifeng's account relayed through it; the design post itself was not read first-hand)Read source