Deep Seek's Super-agent prototype completed half of tasks within 12.4 minutes and 95% within 15.7 minutes in preliminary tests on synthetic workloads of roughly 300 messages containing about 20 work requests2. The team called the results a preliminary reference. Batching under load lowered inference time per Router decision from 6.4 seconds to 1.6 seconds, and without direct routing of follow-ups to Workers, Router decisions rose by about 30%.
Deep Seek published dsh-superagent-prototype on October 31, a separate MIT-licensed repository containing a research prototype of a proactive coding agent built on its open-source Harness runtime3. The company describes the prototype as reference code rather than an official product and warns it is not production-ready.
How the prototype works
The core idea is a two-layer architecture. A resident Router session takes in all human interaction across channels and only dispatches tasks; it does not do the work itself and may not even have a Bash tool. Isolated Worker sessions run tasks concurrently with full tools and report back to the Router instead of talking to users. Each Worker subscribes to the threads it cares about, such as a pull request or a chat topic, so follow-up events go straight to it without passing through the Router.
The system is event-driven: chat messages, emails, review comments, and CI results arrive as triggers, filtered first by deterministic code and then by a model. An authorized request or a team-defined rule can create a work item, which runs in its own agent session and directory before the system responds in the original thread, potentially with a pull request. A web panel surfaces work that requires human attention. The team runs an external Watchdog session that checks whether Workers are making real progress.
In the team's internal user group, an instance silently reproduces reported bugs and speaks up only once it has a simple fix; it sometimes opens a pull request before anyone has tagged it.
ANALYSIS The architecture is event-driven and scoped to repository tasks rather than an open-ended autonomous loop, a design choice that limits the agent's blast radius while still enabling multi-step coding workflows.
Security caveats and open problems
Deep Seek advises users to review plugins and generated commands and to run the software with minimal privileges. The company says its sandboxing and approval controls do not guarantee isolation and that the software has not undergone a security audit. Compatibility-breaking changes may occur.
The accompanying design write-up lists prompt-injection risk when a cloud agent can reach credentials on local devices, agent-to-agent communication, long-term memory, and scaling the Router as open problems. The team separates agent state from execution environments so the agent can stay online in the cloud while reaching a user's own machines.
Harness foundations
The prototype sits atop Deep Seek Harness (often shortened to dsh), which the company released on August 13, 2026 under an MIT license. Harness can be cloned, inspected, modified, and self-hosted; installing a supported version of NodeJS and running npx @deepseek-ai/dsh web opens a local web UI at 127.0.0.1:3080. Its architecture treats model adapters, tools, sessions, sandboxes, storage, agent loops, and the interface as composable plugins running on Cordis, a framework for dynamically composing components and reversing their effects when they are unloaded.
The deepseek-harness repository hit 177,904 GitHub stars within eight days of its August 13 launch and has since climbed past 188,000.
Deep Seek says some ideas in its accompanying Super-agent design article are not implemented in the published code. Deep Seek's open-source agent tooling is separated from its model API business.
The release follows a period of scrutiny for the company. Deep Seek and Moonshot AI relied heavily on output from OpenAI's GPT-6 Astra to train their newest releases, according to prior reporting. Anthropic has separately accused Deep Seek of conducting large-scale distillation attacks against Claude, an allegation Anthropic documented in February 2026.
Harness remains experimental preview software.