On October 31, DeepSeek’s Harness team published dsh-superagent on GitHub — an MIT-licensed coding agent that runs as a persistent background service next to a single GitHub repository. No terminal session required. It monitors GitHub issues and, optionally, Feishu team chats, picks up assigned work, and replies in-thread with a pull request. That makes it the first open-source implementation of what developers have been calling the proactive coding agent model.
Proactive, Not Reactive
Every major coding tool right now — Claude Code, Codex, Cursor — is reactive. You open a terminal, you ask, it does. Close the terminal, the agent stops. dsh-superagent inverts that model. It runs as a long-lived background service. Assign a GitHub issue to the bot account, and the agent picks it up, works it, and comments back with results — whether you’re at your desk or not.
DeepSeek describes this as treating “the ticket as the unit and the session as disposable infrastructure.” The practical result: you can queue up a half-dozen issues and go to lunch.
Three operating modes ship out of the box. Off disables autonomous behavior entirely. Observe (the default) logs what the agent would have done without executing — a smart way to audit before granting real autonomy. Auto gives it the keys.
How the Router–Worker Architecture Works
Under the hood, two components handle the workload. The Router is the always-on process — it watches GitHub events and Feishu messages, matches them against configured rules, and spawns work items. Each work item gets its own Worker: an isolated agent session in a dedicated directory. The Worker executes the task, then posts results back in the original GitHub thread or Feishu conversation.
The model is pluggable. dsh-superagent uses whatever model you configure in your DeepSeek Harness profile — not locked to DeepSeek’s own API. Claude, GPT-4o, or a local model all work, provided the harness supports the provider.
Read the Security Warning Before You Install
DeepSeek is unusually candid in its own documentation: dsh-superagent “runs with full operating-system user access and should be isolated before use.” That’s not boilerplate. Every Worker session runs in danger-full-access mode — hard-coded, not configurable. Workers can execute programs by arbitrary path and read your home directory, without approval prompts.
There’s also CVE-2026-82533 (CVSS 9.4), a critical vulnerability disclosed in September in DeepSeek Harness: a sandboxed agent could call the harness’s unauthenticated internal API via a single curl command and escalate to full access with all approval prompts disabled. DeepSeek’s own docs add: “Sandboxing does not guarantee isolation or prevent damage.”
The mitigation is non-negotiable: run the agent under a dedicated OS account or a disposable VM. Keep no personal SSH keys, cloud credentials, or production secrets in that environment. The dedicated GitHub identity should have access only to the target repository.
Setup Requirements
Getting dsh-superagent running requires Node.js 22.19+, pnpm, git 2.32+, tmux, Python 3.8+, GitHub CLI 2.48+, DeepSeek Harness, and a dedicated GitHub bot account. That’s a non-trivial list.
DeepSeek’s recommended setup method: clone the repository, open it in Claude Code (yes, their competitor), and tell the agent “Read SETUP.md in this repository and set up dsh-superagent with me.” It walks through the installation one question at a time. Using a competitor’s agent to set up your competitor’s agent is either pragmatic or ironic, depending on your mood.
dsh-superagent vs. Claude Code and Codex
dsh-superagent isn’t competing on benchmark scores — no SWE-bench numbers exist for this prototype. The architectural bet is different: background execution beats on-demand sessions for teams running multiple independent tickets in parallel.
| Feature | dsh-superagent | Claude Code / Codex |
|---|---|---|
| Execution model | Proactive background service | Reactive terminal session |
| Cost | Free (MIT, self-hosted) | API pricing |
| Security | Full OS access, needs isolation | Sandboxed per session |
| Maturity | Research prototype | Production-ready |
| Setup | High complexity | Low (minutes) |
Bottom Line
The ticket-centric, background-service model is the right architectural direction for coding agents in team environments — dsh-superagent demonstrates that convincingly. The security posture makes it a research experiment today, not a production deployment. Run it in an isolated environment, study the design, and adapt the MIT-licensed code to your needs. The open-source, model-agnostic approach is the right foundation for where coding agents are headed. Just don’t run the workers as your own user.
