Chinese artificial intelligence company DeepSeek released an experimental coding agent prototype called dsh-superagent, introducing a two-layer router-worker architecture designed to handle parallel tasks and reduce human supervision bottlenecks.
DeepSeek Harness Team Releases dsh-superagent Prototype
The DeepSeek Harness team published a technical design post titled Super-agent Design Exploration: Toward a More Autonomous Agent System
via its official WeChat account, detailing an architecture aimed at multi-task, team-collaboration scenarios according to the reporting outlet. Alongside the design document, the team released an MIT-licensed research prototype bundle named dsh-superagent version 0.1.0 on GitHub.
The tool functions as a proactive coding agent for a single GitHub repository. Set up within a local environment via a designated DeepSeek Harness profile, it depends on Node.js 22.19+ or 24+, git 2.32+, Python 3.8+, tmux, and GitHub CLI 2.48+. While an internal Super-agent instance has already been deployed into DeepSeek’s daily research and development work—handling bug reproduction, code reviews, and framework iterations—the published software remains an engineering exploration rather than a formal, standalone commercial product. As noted in the published documentation, the prototype is packaged as a bundle for DeepSeek Harness that listens to GitHub and, optionally, Feishu, turning allowed requests into parallel work items where each runs in its own agent session.
Router-Worker Architecture and Parallel Session Control
Global Times reported that an individual can actively track only three to five active sessions simultaneously.
To overcome this, the research team structured the system around two layers: a long-running Router that handles interaction and dispatch, and isolated Worker sessions that execute each Ticket. A router dispatches work, while isolated worker sessions execute each ticket.
- Router: A long-running component responsible for managing user communication, maintaining shared context, and dispatching tasks using lightweight models.
- Worker: Isolated sessions running independently to execute assigned work items, known as Tickets, which contain an immutable mandate.
The system incorporates a subscription mechanism where subsequent comments on a pull request route directly to the assigned Worker without requiring a new dispatcher evaluation. Proactive triggers include group-chat messages, emails, calendar reminders, GitHub comments, and continuous integration test results. An authorization module enforces three accept modes: off, observe (the default, which records accepted actions without executing them), and auto. Furthermore, tasks can resume when a temporarily offline device comes back online because only a lightweight daemon needs to be deployed locally to receive cloud commands and return results. Functioning much like an AI supervisor, a separate Watchdog mechanism monitors whether tasks progress normally or get stuck in unproductive loops, intervening to attempt fixes whenever issues are detected.
Simulation Performance Metrics and Resource Costs
Under noisy conditions, the Super-agent completed tasks significantly faster than conventional single-session baselines using neutral prompts.
| Performance Metric | Super-Agent Architecture | Conventional Single-Session Baseline |
|---|---|---|
| Task completion time (p50) | 12.4 minutes | Not disclosed |
| Task completion time (p95) | 15.7 minutes | 29.5 minutes |
| Cache-read tokens (adjusted) | 31.77 million | 23.23 million |
| Router decision volume change | Up approximately 30% | Baseline reference |
The efficiency gains introduce higher token consumption. Because the system maintains multiple independent sessions, cache-read tokens increased by about 37% after removing event pass-through, while Router decisions rose by roughly 30%. When event arrival rates escalated from one message per minute to 40 per minute, batch processing maintained throughput, pushing single-request latency from 2.5 seconds to 3.9 seconds, with batch size growing from 1.0 to 2.8.
Industry Analyst Perspective and Competitive Context
The unveiling arrives amid intensifying rivalry between Chinese artificial intelligence developers and US counterparts such as OpenAI and Anthropic, whose Claude Code and Codex platforms similarly automate software engineering workflows according to engineering analysts. Observers note that industry competition is expanding beyond raw model parameters into workflow scheduling and system reliability.
AI agents are evolving beyond simple interfaces for calling large-language AI models into more sophisticated systems capable of handling task scheduling, storage, isolation and fault recovery.
Ma Jihua, veteran telecom and technology industry analyst, via Global Times
Ma added that task success rates, failure recovery, and context management costs have become primary competitive differentiators as domestic firms push AI agents deeper into business operations alongside comparable enterprise tooling from Tencent and Alibaba Cloud.