A Coding Agent That Acts Before You Ask

Every major coding agent shipped in 2026 — Claude Code, Cursor, Copilot — shares one trait: they are reactive. You open a terminal or IDE, invoke the tool, and the agent responds. DeepSeek’s Harness team published a prototype that flips this pattern. The agent watches your repository and acts when conditions are met, without a developer typing a command.

The project is called dsh-superagent, released under the MIT licence on October 31, 2026. It reached roughly 3,700 GitHub stars within 24 hours, according to Eneralabs. A write-up titled “Super-agent design exploration” by Yujia Qiao, published on the same day, describes the design philosophy and architecture.

The stated goal, per the documentation, is to “take as little human attention as possible, and give as much value as possible.” That means the agent confirms a task once, stays quiet during execution, and reports back only with a result, a question, or a failure.

How the Router-Worker Architecture Works

The prototype separates concerns into two layers, as described by CellCog’s technical analysis.

A resident Router runs continuously at low cost. It subscribes to GitHub events and, optionally, Feishu group chats. Every incoming message passes through a code filter first; only what passes goes to the model for a decision. When the Router identifies a legitimate work request from an authorised user, it creates a Ticket.

Each Ticket spawns an isolated Worker session with its own directory and full tool access. Workers run in parallel — if three engineers file separate requests within minutes of each other, all three Workers run simultaneously. Each Worker executes the assigned task and reports back through an Engine layer that routes results to the original conversation thread. A separate watchdog session monitors each Worker for real progress.

At 40 triggers per minute in simulation, Router inference time fell from 6.4 to 1.6 seconds per decision. The parallel architecture holds under load — in simulation, at least.

What the Benchmarks Actually Measure

The benchmarks DeepSeek published with the release are throughput and scheduling metrics, not coding quality scores. Each simulated run used about 300 chat and CI messages, with roughly 20 actual work requests. Critically, coding was replaced with a tool that only consumes time — the agent never wrote or reviewed real code during these tests.

Against a single-session baseline, the Super-agent reached p50 completion at 12.4 minutes and p95 at 15.7 minutes. The baseline’s p95 was 29.5 minutes — a 1.9× improvement at the tail, per Eneralabs. These figures show efficient scheduling. They say nothing about whether the agent produces correct, secure, or review-ready code.

The prototype’s own README is direct about this gap: the released code is not identical to DeepSeek’s internal version, some designs described in the write-up are future directions rather than implemented features, and the code “may contain bugs and inconsistent style,” per AI Weekly. Also worth noting: the prototype runs commands as the operator’s user without per-command approval. DeepSeek explicitly recommends treating it as research code.

Reactive vs Proactive: Where the Landscape Stands

The table below compares the main coding agent options as of November 1, 2026. Information is from vendor documentation and the sources linked in this article.

One important distinction: Claude Duet, ChatGPT Dots and Meta Muse are also proactive — but they are conversational productivity assistants monitoring calendars and messages. The dsh-superagent’s unit of work is a code change filed as a pull request. Different category.

Running Cost: A Back-of-the-Envelope Estimate

The MIT licence means no platform fee. Your cost depends on which model you wire to the Workers. The prototype does not specify one. If a team uses DeepSeek V4.1-Flash — benchmarked at 77% on DeepSWE — here is what one typical coding task might cost, using the API prices DeepSeek published in July 2026 (input $0.14/MTok, output $0.28/MTok, per DeepSeek’s GA announcement; check current pricing — rates were updated in September 2026):

• 10,000 input tokens (task context + relevant files): $0.0014
• 2,000 output tokens (code changes + PR description): $0.00056
• Estimated per task: ~$0.002

Agent loops consume more tokens than single queries. A task involving multiple reasoning steps, tool calls, and error recovery on a medium-sized repo is likely 5–10× this estimate — call it $0.01–$0.02 per task. At $0.02, 1,000 automated tasks cost $20. That is below per-seat SaaS pricing at comparable volume, assuming the work is actually done correctly — which the benchmarks do not yet tell us.

For Swiss & EU teams

The MIT licence lets you run the Router and Workers on infrastructure you control — on-premises or an EU-region cloud provider — so no code, no context, and no conversation data needs to leave your network. That is the simplest path to FADP compliance for a coding agent. The prototype integrates with the DeepSeek Harness platform out of the box, but the code is open enough to adapt. There is no documented EU self-hosting guide yet; budget engineering time for the setup. The Observe mode (logs decisions without acting) is the right starting point for any team that wants to evaluate the Router’s judgment before granting write permissions to a live repository.

What to Do on Monday

1. Read the README before cloning. The authors are clear: it runs as your user without per-command approval and the code has known bugs. Understand the permission model before connecting it to anything.
2. Start in Observe mode. Run shadow trials for one to two weeks on a real repository. Log every decision the Router would have made. Only promote to Auto once you trust its judgment on your actual work patterns.
3. Use a throwaway repo for the first real run. A sandbox repo with sample issues lets you see the full Router → Ticket → Worker → PR flow without risk to production code.
4. Log token usage from day one. Add usage tracking to each Worker session so you know actual cost per merged PR before you scale. The estimate above is rough.
5. Wait for a coding quality benchmark before calling it production-ready. The routing speed numbers are real. The code quality is unverified. No independent SWE-bench or Terminal-Bench run exists yet. When one drops, we will cover it.

Further Reading

• DeepSeek dsh-superagent: Proactive Coding Agent Prototype (AICoder) — release overview with the Router-Worker architecture write-up
• AI Coding Agents in 2026: Benchmarks, Costs, and How to Choose — our pillar comparing 7 agents on Terminal-Bench 4.0 and cost-per-task math
• DeepSeek Super-Agent: Autonomous Coding for Enterprise Teams (Eneralabs) — enterprise framing with the GitHub stars count and the DeepSeek funding context

Your turn: Is your team already running any proactive or background coding agent mode — Cursor’s background agents, Claude Code’s dynamic workflows — and has it actually reduced your own interruptions? Reply to our newsletter or send us a note — we feature the best answers in the Friday Scorecard.

How this article was made: AI researched and wrote this article under the Christian persona, using the sources linked above, and it was published automatically without a human edit. Editorial guidelines are set by Adi. Spotted an error? Tell us and we will correct it.