DeepSeek open-sources a proactive coding agent: dsh-superagent-prototype watches group chats, dispatches Router and Worker sessions to fix bugs

On October 31, DeepSeek’s Harness team published a design article, “Exploring Super-agent Design: Towards a More Autonomous Agent System,” and the same night pushed the matching prototype to GitHub as dsh-superagent-prototype under MIT. The thing is a proactive coding agent for one GitHub repository, with optional Feishu group-chat intake: when a team member files a bug or asks for a change, the agent opens a work item, does the job in its own isolated session, and reports back into the original conversation — usually with a pull request. The GitHub API shows the repo was created at 09:32 UTC on 2026-10-31 and had drawn about 6,900 stars by November 2. “Harness” is DeepSeek’s name for the scaffolding around a model — tools, sessions, permissions — and the team, set up earlier this year, works on exactly that layer.

Key points

• The article lays out a three-layer Router-Worker-Watchdog architecture; the open-source prototype implements only Router and Worker. The Watchdog and an asynchronous ReAct loop are listed as future directions, not code.
• Proactive work is gated by an accept mode: the default, observe, only records what the agent would have accepted. Real autonomous action requires switching to auto. Explicit requests that @ the agent run in every mode.
• In the team’s own simulations, the super-agent finished requests at a p95 of 15.7 minutes against 29.5 minutes for a single-session baseline — from synthetic messages and simulated tools, with no test of real coding quality.

Background

The motivation starts from an observation: the DSH team argues a person can track only about 3 to 5 active agent sessions at once, so the bottleneck is human attention, not model capability. Handing agents more initiative means answering when they should stay quiet. TMTPost’s Chinese write-up compares the approach with Anthropic’s Claude Duet preview, which sends a morning briefing and checks mail and calendar every few hours but asks approval before sending, paying or deleting. dsh-superagent takes a third path: rule-gated proactivity, encoded in the repo’s philosophy doc — “take as little human attention as possible, and give as much value as possible.”

Proactive agents are clearly a direction the whole industry is moving in: GitHub’s own Octoverse 2026 found that agents now draft more than 40% of pull requests, and LangChain open-sourced the harness behind its design agent last week. What separates this release is that it turns the hard questions — how to gate, how to dispatch, how to reproduce a bug — into a reference implementation anyone can read.

The facts

The workflow runs like this: a Source (GitHub or Feishu) converts events into Triggers, which pass a two-layer Eval — hard rules first to drop bot messages and out-of-scope requests, then a model judgment on what remains. Passing both opens a Ticket with a Mandate that states its authorized scope, and hands it to a Worker in an isolated session. Shell commands that run longer than 5 seconds move to the background so the Worker keeps receiving messages; follow-up comments on its pull request route back to the same Worker through Subscriptions, no @ needed. A web panel shows only what waits on a person.

All benchmarks are self-reported, from simulations with synthetic group-chat messages and simulated tools — roughly 300 messages, about 20 real requests, 5 runs per scheme:

Metric	Super-agent	Single-session baseline
p50 completion	12.4 min	not reported
p95 completion	15.7 min	29.5 min
Subagent calls in 5 runs	—	0

A Router load test pushed the arrival rate from 1 to 40 messages per minute: batch size grew from 1.0 to 2.8 and per-request latency from 2.5 to 3.9 seconds. In an ablation that removed direct event pass-through, Router decisions rose about 30% and cache-read tokens rose from 23.23M to 31.77M — the batching saves attention by spending tokens.

MarkTechPost’s mechanism comparison against rival tools: Claude Code Action triggers on an @claude mention, Copilot’s cloud agent on issue assignment and automation rules, and OpenAI Codex responds to @Codex in Slack. None of the three discloses per-task isolated sessions, while dsh-superagent runs one per Ticket and ships under MIT.

What others say

Outside coverage has been measured. Digital Trends frames the release as a move against Claude Code and Codex — the headline talks about stealing their lunch — but says plainly that a prototype is not proof DeepSeek has overtaken anyone, and it is no reason to uninstall either rival yet. MarkTechPost highlights two boundaries: the observe default and rule-based authorization for proactive work, plus the fact that the agent runs with full OS-user access, so operators should isolate it before deployment. TMTPost’s reading centers on the restraint philosophy: the design problem is how an agent can act on its own without nagging, not how it gets smarter. Community traction is still thin — four Hacker News submissions, the top one at 11 points with zero comments.

Our take

The value here is the design documents, not the code. The README states up front that this is research code, not an official DeepSeek product; the public version does not match the internal implementation the article describes; and no third party has tested its coding quality. For engineers building an internal agent platform, though, it is a runnable architecture reference — Ticket-based authorization, two-layer filtering and subscription routing can each be lifted on their own. Keep three boundaries in mind: the Watchdog does not exist yet; “fixes bugs and opens PRs without being asked” describes the auto mode as a design scenario, while the default only observes; and the benchmarks are simulated, so treat them as direction, not performance. It also runs with your full user permissions — sandbox it first. The open question nobody has answered: how far the public version sits from whatever DeepSeek runs internally.

How to try it

The path runs through SETUP.md in the repo: clone deepseek-ai/dsh-superagent-prototype, then ask your coding agent to read SETUP.md and set it up with you — it asks one question per message, shows the full plan, and executes only after you confirm. Prerequisites: Node.js 22.19+ or 24+, pnpm, git 2.32+, tmux, Python 3.8+, GitHub CLI 2.48+ and the DSH CLI, which needs a model provider configured. Keep the default observe mode for the first few days, check whether the panel’s “would accept” records match your judgment, and only then consider switching to auto.