Skip to content
Agents tracked: 492 Releases (7d): 545 Agent status: 1 with issuesSince last limit reset: Claude 9d · Codex 7d For AI agents: Markdown Updated Nov 20, 2026

✓ Checked

DeepSeek Harness's Super-agent Design and Its Open Prototype

A long X article from the DeepSeek Harness team describes an always-on agent: one router session, a worker per task, event triggers and a watchdog. We checked it against DeepSeek's MIT prototype on GitHub: what ships, what is still design, and what the numbers mean.

On November 1, 2026 the X account DeepSeek Harness (@DeepSeekHarness) published a long X Article, "Exploring Super-agent Design: Towards a More Autonomous Agent System". It is a machine translation of a Chinese original. It describes how the team behind DeepSeek Harness runs a single agent that takes work from group chats and GitHub at the same time, so that nobody has to juggle separate chat sessions. The article says at the top that it is about architectural ideas, not an introduction to features DeepSeek Harness already has.

We checked who published it, what code exists and how the design maps onto that code. In short: the design is backed by a real MIT-licensed repository in DeepSeek's GitHub organization, dsh-superagent-prototype. That repository calls itself a research prototype, not an official DeepSeek product, and the experiment numbers in the article come from synthetic workloads that the authors themselves call indicative only.

Who published it

We could not confirm the X account from an official page. DeepSeek's Harness page, the deepseek-harness README and the deepseek-ai organization on GitHub do not link to an account called @DeepSeekHarness; the social links on the Harness page point to @deepseek_ai, and the README sends people to GitHub Discussions and a Discord server.

The article itself is easier to place. The README of deepseek-ai/dsh-superagent-prototype names it by its full title and says the published code differs from both the team's internal version and the implementation the article describes. The article's credits name the DeepSeek Harness team and one engineer on it as the person who built the system and wrote the text. So the content comes from the DeepSeek Harness team; the account is, as far as official pages show, unverified.

DeepSeek Harness itself is DeepSeek's open-source agent harness, built so that every capability is a plugin. DeepSeek's page says it is in public preview worldwide; its README calls it a developer preview and warns of compatibility-breaking changes. DeepSeek's page says it runs as a desktop app or launches its web UI from code; npx @deepseek-ai/dsh web starts the web UI on 127.0.0.1:3080.

The design in one picture

The article's starting point is attention. Coding agents need less and less supervision on a single task, but a person can follow only 3 to 5 active sessions at once, as reported. Its answer is to hide sessions from the user altogether: one long-lived agent that people reach in chat or on GitHub, which splits the work underneath.

EventsGroup chat messages, emails, calendar reminders, review comments, CI results
→
FilterCode rules first (source, group, allowlists), then a model for what is left
→
RouterOne long-lived session: talks to people, dispatches tasks, keeps memory; no Bash tool
→
WorkersOne session per task with full tools, subscribed to that task's Threads
Outside the loop: a Watchdog session on a schedule checks failures and whether each Worker is still making progress.
Our drawing of the Super-agent architecture as the article describes it. A Thread is the stream of events on one pull request, chat topic or email chain.

The five parts, in our words:

  • Two levels. The Router handles every human exchange and only dispatches; Workers each own one task, never talk to people directly and report back to the Router. The article says a lighter model is usually enough for the Router, while Workers need the strongest one.
  • Events instead of prompts. Most incoming messages need no action, so cheap code checks run before any model call (the article borrows "fast path / slow path" from systems design). Once a Worker owns a Thread, follow-up events go straight to it ("pass-through") instead of through the Router. Context is injected sparingly with IDs and short summaries, and the model follows those IDs to recall more ("Chain of Recall"). The agent may also start work ahead of time, for example silently trying to reproduce a reported bug and speaking up only if it finds a simple fix.
  • Full duplex. Long tool calls become start() and stop() so the session can take new instructions while work runs in the background. As reported, DeepSeek Harness already implements this, but it is not on by default.
  • Channel manners. Short replies in chat (the team uses a cap of 100 words per reply), full write-ups on GitHub pull requests and in email.
  • Compute, state and machines apart. The agent loop runs in stateless cloud containers with sessions in a database; a small daemon on the user's own computer runs commands ("Bring Your Own Machine"). The article compares this with an Anthropic engineering post on the same idea.

What the experiments show, as reported

The article ends with two small experiments. In the first, four designs ran on a stream of about 300 group chat and GitHub CI messages, about 20 of them real work requests, 5 runs each. Coding was simulated by a tool that only takes time, so the test measures scheduling, not code.

Super-agent, p50
12.4 min
Super-agent, p95
15.7 min
Single session, neutral prompt, p95
29.5 min
Task completion time from a work request to its result, as reported in the article (section 7.1). The article gives no single-session p50 in its text.

With the neutral prompt, the single session never handed work to a subagent in any of the 5 runs; with a prompt that pushed delegation it did so in 4 of 5. By our calculation the neutral single session's p95 was about 1.9 times the Super-agent's. Turning off pass-through left completion times about the same, but the Router made about 30% more decisions and cache-read tokens rose from 23.2M to 31.8M per run, as reported.

The second test pushed events at the Router alone, from about 1 to about 40 per minute:

Measure, as reported At about 1 per minute At about 40 per minute
Average events merged per request 1.0 2.8
Time per model request 2.5 s 3.9 s
Time per decision 6.4 s 1.6 s

Events that pile up during one turn are merged into the next model call, so by our calculation the time per decision fell to a quarter while each request took only a little longer. That is the case for keeping one Router session rather than splitting it early.

Keep the limits in mind. The authors call these results indicative only; the workloads are synthetic, the article does not name the models, and costs are given in cache-read tokens, not money. The Super-agent design also uses more tokens than a single session, which the article states openly: it trades tokens for shorter, steadier waits.

What the prototype actually ships

dsh-superagent-prototype is an MIT-licensed TypeScript repository created on October 31, 2026, the day before the article; it had 19,574 stars when we read GitHub's API on November 6, 2026. It is a bundle for DeepSeek Harness: a proactive coding agent for one GitHub repository that can also listen to Feishu chats. Each accepted request becomes a Ticket with its own agent session and directory, the answer goes back to the original conversation (often as a pull request), and a web panel lists what needs a person.

Most of the article's vocabulary is in its docs, sometimes under another name:

In the article In the prototype's docs
Router layer The "main Session" (orchestrator in code), which only routes and keeps routing notes short
Worker An Executor working on one Ticket in its own directory
Trigger, Thread The same names; a Source (GitHub or Feishu) produces Triggers
Two-stage filtering Eval: a deterministic gate, then the model's assessment
Pass-through Subscriptions: a many-to-many link between Tickets and Threads
Acting ahead Team-written rules with an accept mode; the default observe only records what the agent would have done
100-word cap Chat messages are one sentence by default, 100 words at most
Watchdog An hourly check, scheduled from your other DeepSeek Harness profile, that restarts the service when it is down
Cloud split and BYOM Not part of setup: it runs as a separate DeepSeek Harness profile on your machine under tmux, with the panel on 127.0.0.1

The README is direct about the gap: some designs in the article are future directions and are not in this prototype. The hourly check looks after the service; the docs we read do not describe a model judging each Executor's progress the way the article's Watchdog does.

What the post gets right

  • The code exists and is DeepSeek's. The prototype sits in the deepseek-ai organization under MIT, and its docs use the article's terms.
  • The details match. The 100-word cap, Thread subscriptions and the code-then-model filter are all written into the prototype's prompts and domain model.
  • It is honest about itself. It says the experiments are indicative, lists open problems, and says plainly that the plugin is a prototype, not an official DeepSeek product.

What it leaves out

  • Who the account is. No official DeepSeek page we found links to @DeepSeekHarness.
  • How much is built. The cloud split, the daemon on your own machine, a model-driven Watchdog and background work by default are design, not the code you can download.
  • What it takes to run. Setup needs Node.js 22.19+ or 24+, pnpm, git 2.32+, tmux, Python 3.8+, GitHub CLI 2.48+ and a model provider configured in DeepSeek Harness, plus a separate GitHub account for the agent.
  • The risk. The article's own open-problems section says a hijacked cloud loop could take over every connected device and that credentials on a machine can leak into the model's context. The prototype's answer for now is narrow scope: a dedicated agent account, a short allowed list and observe mode until the team trusts it.

AgentGid data

On AgentGid, DeepSeek Harness had 272,983 GitHub stars and 812,609 weekly package downloads (+25.0% over 30 days) as of 2026-11-05, with a Gid Score of 63.8: #17 of all tracked agents and #3 by Gid Score among general-purpose agents, after OpenClaw and Hermes Agent. Gid Score measures adoption and attention, not quality. Our guide to Google's Harness² and the harness engineering course covers the other side of the same topic: what goes around a single agent session.

How to try it with a coding agent

The README does not ask you to run setup by hand. You open the checkout in your coding agent (Claude Code or DeepSeek Harness itself) and ask it to read SETUP.md and set things up with you. The agent asks one question at a time (sources, repository, agent account, who may ask for work, checkout, commit author, auto-merge), shows one plan, and only then installs. Two defaults are worth keeping: a dedicated GitHub account for the agent, never your own, and observe mode for anything the agent was not asked to do. Start on a test repository and read what it would have done before you switch rules on.

The verdict

  • A serious design note with real code behind it. The article is the clearest public write-up we have seen of an always-on agent that works across chat and GitHub, and the prototype implements its core: one routing session, per-task workers and Thread subscriptions.
  • Not a product, and not all built. The repository says it is research code, and the cloud and Watchdog parts of the article are not in it.
  • Read the numbers as a sketch. Synthetic messages, a timer instead of coding, no model names: useful to see the trade-offs, not to compare tools.
  • Worth reading if you run agents for a team. The ideas (code filters before models, one router, short chat replies) apply to any harness, including Claude Code with subagents.

Facts about the prototype come from its README, docs and GitHub's API on November 6, 2026; facts about DeepSeek Harness from its README and DeepSeek's Harness page; the experiment figures are as reported in the article. Ratios are by our calculation. We did not run the prototype or check the account beyond the official pages named here.

Was this useful?

Agents mentioned in this guide

# Agent Gid Score Downloads 7d 7d 30d Stars Latest release
10 67 658K down 22.5% up 59.3% 284K+685/day —

Live table, updated daily (figures as of Nov 20, 2026); numbers in the text above are as of the dates given there. "#" is the rank among products of the same kind by Gid Score, our 0-100 composite of adoption, community, attention and momentum (how it is calculated); a dash means no public data.