DeepSeek dsh-superagent: Router-Worker Agent Hits p95 of 15.7 Min

DeepSeek's Harness team has released dsh-superagent, an MIT-licensed TypeScript bundle that converts GitHub issues, pull request comments, and optional Feishu messages into parallel, isolated coding-agent sessions. The project is explicitly not an official DeepSeek product — self-described as research code built largely by coding agents — but it embodies a concrete architectural claim: that human attention, not model capability, is the real bottleneck in multi-task agent workflows, and that a router-worker split can cut p95 completion time by nearly half on simulated workloads.

The repository had accumulated 7.1k stars and 712 forks at time of writing. It requires Node.js 22.19+ or 24+, git 2.32+, Python 3.8+, tmux, and GitHub CLI 2.48+. Setup is guided interactively by a coding agent following SETUP.md.

Router-worker architecture

The system separates two concerns into two persistent layers. A long-running Router handles all inbound events under the dsh-superagent-router preset, using a reduced tool set to evaluate and dispatch work. Each dispatched Ticket spawns a Worker running the dsh-superagent-worker preset inside its own session and directory. Workers never share state; shell commands longer than 5 seconds move to the background automatically, keeping the Worker's message queue live. Background work is allocated up to 30 minutes per turn.

The event pipeline is: Source → Trigger → Eval → Ticket → Channel reply. A Trigger is a single immutable event; submitting the same Trigger twice still creates exactly one Ticket. Each Ticket carries an immutable Mandate recording why it was authorized to proceed. Subscriptions route later pull request comments back to the originating Worker without requiring a fresh @-mention.

Under load, MarkTechPost reports that as message arrival rates rose from roughly 1 to 40 per minute, the Router's average batch size grew from 1.0 to 2.8, while per-request latency rose from 2.5 to 3.9 seconds — figures DeepSeek reports from its own simulation.

Authorization is tiered across three modes. In off, only explicit direct requests run. In observe — the default — inferred work is assessed and logged as "would accept" without starting any Ticket. In auto, a positive assessment triggers a Ticket immediately, subject to at least one team-written rule permitting it. This graduated exposure model is relevant to AI agent security on shared surfaces, where uncontrolled proactive action is a primary attack surface.

Benchmark results

DeepSeek's simulation ran approximately 300 messages per trial, of which roughly 20 were real work requests, repeated 5 times per scheme. All figures are vendor-reported, as relayed by MarkTechPost and ifeng from DeepSeek's own design essay.

The super-agent configuration achieved a p50 task completion time of 12.4 minutes and a p95 of 15.7 minutes. The single-session-plus-subagents baseline under a neutral prompt reached a p95 of 29.5 minutes — roughly 1.9x slower, per MarkTechPost. That baseline called a subagent zero times across all five runs, meaning the comparison is essentially between parallel dispatch and fully sequential handling. An ablation removing direct event pass-through caused Router decisions to increase by approximately 30% and cache-read tokens to rise from 23.23 million to 31.77 million, a roughly 37% increase.

Comparison with other GitHub coding agents

AI Mastery analysis

The benchmark's most revealing detail is not the 15.7-minute p95 but the baseline's zero subagent calls across five runs. A neutral prompt failed to invoke parallelism at all, which means the speedup is not about model capability — it is about architecture forcing parallel dispatch regardless of what any single model session decides to do. This is consistent with a pattern tracked across multiple infrastructure releases: pipeline architecture, not better models, is where 2026 AI gains are being captured.

Two practical constraints warrant attention before adoption. Both presets run as danger-full-access under the operator's OS user. DeepSeek recommends a dedicated OS account for isolation, but that recommendation is advisory, not enforced — teams should treat OS-level isolation as a hard prerequisite. The simulation measured scheduling throughput across synthetic messages with simulated tools; it did not evaluate code correctness or pull request acceptance rates. Operators should run evals against their own repositories before drawing conclusions about end-to-end productivity. GitHub Enterprise Server is unsupported, and Feishu is the only non-GitHub chat source shipped, which narrows the addressable deployment surface for enterprise teams on hosted Git infrastructure.

The observe default is a sound choice for a research prototype with full OS access, and the immutable Mandate per Ticket provides an audit trail that most agent frameworks omit. Features described in the accompanying design essay — an external Watchdog, storage and compute separation, and an asynchronous ReAct loop — are listed as future directions not yet implemented. The prototype is a proof of concept for the router-worker pattern, not a feature-complete agent platform.

DeepSeek is publishing runnable architecture alongside its research writing, giving engineers something to instrument and test rather than only a paper to cite. The router-worker split is straightforward enough to port to other orchestration stacks, and the MIT license removes any barrier to doing so.

Sources

• Primary: GitHub - deepseek-ai/dsh-superagent-prototype: Towards a More Autonomous Agent System · GitHub
• Reported by: MarkTechPost: DeepSeek Open-Sources dsh-superagent, a Router-Worker Coding Agent