On Oct 31, deepseek-ai open-sourced dsh-superagent-prototype on GitHub under the MIT license: a proactive coding agent that runs as a DeepSeek Harness bundle, lives on one GitHub repository and can optionally listen to Feishu. When an allowed person asks for work, it opens a work item, works in its own session and directory, and answers in the original conversation, for example with a pull request. The README states it is a research prototype, not an official DeepSeek product. The companion article, Exploring Super-agent Design, proposes a two-layer Router-Worker architecture; the team test, as reproduced by media, reports a p95 completion time of 15.7 minutes versus 29.5 minutes for a single-session setup.
Key Takeaways
- ✓Open source: deepseek-ai/dsh-superagent-prototype went public on Oct 31 under MIT, written in TypeScript, with about 1,800 stars at the time of writing
- ✓Scope: each deployment serves 1 GitHub repository, with optional Feishu; it acts only when an allowed person addresses it or a team-written rule allows it
- ✓Delivery: by default it only allows auto-merge for an approved pull request opened for a ticket with fewer than 200 changed lines, and never enables auto-merge as the agent account
- ✓Team test (as reproduced by media, vendor-reported): in a simulation with about 300 messages and 20 requests, p50 / p95 completion was 12.4 / 15.7 minutes versus 20.4 / 29.5 for a single session with a subagent tool
- ✓Risk: worker sessions run commands without approval as the operator's OS user, and the README asks users to review and adapt the code first
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background and pain point
Most coding agents today work one session per task, so developers end up watching 3 to 5 sessions, switching between them and re-explaining context. The bottleneck has moved from model ability to human attention. The DeepSeek Harness team answers with a resident Super-agent whose principle is on the repo front page: take as little human attention as possible and give as much value as possible.
How it works
According to the team article, a resident Router layer handles human interaction, global context and dispatch without doing the work itself, while several context-isolated Worker sessions run tasks in parallel and subscribe directly to later pull request comments. External events are filtered first by program rules, then judged by a model. The article also proposes speculative execution, an asynchronous ReAct loop, separation of compute, state and execution environment, and a Watchdog that catches stuck or drifting work. The README says the open-source prototype differs from the internal version and some designs in the article are not implemented yet.
Benchmarks and measured results
There is no public third-party benchmark. Per the article figures reproduced by media (team test, vendor-reported), in a simulation with about 300 messages and 20 requests the Super-agent reached p50 / p95 completion times of 12.4 / 15.7 minutes, versus 20.4 / 29.5 minutes for a single session with a subagent tool. Without the direct delivery mechanism, cache-read tokens rose from 23.23M to 31.77M.
Developer takeaways
It needs Node.js 22.19+, pnpm, git, tmux, the GitHub CLI and DeepSeek Harness, and the team recommends letting your own coding agent read SETUP.md and configure it one question at a time. Workers run commands without approval as the operator, so use an isolated machine, start with a short allowed list, and require approvals in branch protection.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.