DeepSeek’s Harness team put a new coding agent called Super-Agent on GitHub on October 31, free, open-source, MIT-licensed — and according to Bloomberg’s November 1 report, aimed squarely at Anthropic’s Claude Code. South China Morning Post’s coverage frames the release as the first concrete public output from a team DeepSeek staffed up earlier in 2026 with a single mandate: benchmark Claude Code and build something that competes with it.

Super-Agent isn’t trying to out-code Claude Code on a hard refactor. It’s trying to redefine what “asking for help” even means — by not waiting to be asked.

Quick Summary: What Happened

Detail	Info
What shipped	dsh-superagent-prototype — open-source, MIT-licensed, built for DeepSeek Harness
Release date	October 31, 2026; covered by Bloomberg and SCMP on November 1-2
What makes it different	Monitors GitHub and Feishu team chats, flags coding problems, and proposes fixes without being summoned
DeepSeek’s own label	”Research prototype for reference only, not an official DeepSeek product”
Built by	DeepSeek’s Harness team, formed earlier in 2026 specifically to compete with US agent offerings
Direct competitive target	Claude Code, per Bloomberg’s framing
Cost	$0 — self-hosted, MIT license

Bottom line: DeepSeek is giving away, for free and under an MIT license, a proactive agent architecture Claude Code and Codex don’t offer in this form — while labeling its own release a prototype, not a finished rival.

What Actually Happened

Per the GitHub repository and Help Net Security’s coverage, dsh-superagent is packaged as a bundle for DeepSeek Harness — the agent runtime DeepSeek has been building since early 2026 — and scoped to a single GitHub repository at a time. It watches two channels: GitHub activity and, optionally, Feishu, the ByteDance-made workplace chat app most Chinese dev teams already run on. When an authorized team member flags a problem — a bug report in a thread, a comment on an issue — Super-Agent opens its own session, does the work, and replies in the original thread, often with a pull request attached.

That’s the part that separates it from basically every coding assistant on the market right now, Claude Code included. Request-response tools wait for an explicit prompt in a dedicated session. Super-Agent treats the team’s existing chat as the interface. No new window to open, no copy-pasting a stack trace into a sidebar. You complain about a bug in the channel where your team already complains about bugs, and — if the agent’s rules allow it — something starts working on it.

DeepSeek is careful not to oversell this. The repo describes it, in writing, as a “research prototype for reference only, not an official DeepSeek product,” and notes plainly that the code “grew out of our team’s day-to-day development work” and isn’t production-ready. Coverage from SiliconReport and CCTest both use a sharper phrase for the same idea: DeepSeek is positioning the release as an architectural exploration rather than a finished product — a way of showing the field what a team-chat-native agent architecture could look like, not a thing to bet a production pipeline on yet.

How Does DeepSeek’s Super-Agent Actually Work?

1. Event filtering first. Programmatic rules strip out irrelevant chatter before anything reaches a model — most of what happens in a dev channel is noise, and Super-Agent doesn’t burn inference on it.
2. A model judges what’s left. For events that survive the filter, a model decides whether the message actually requires action.
3. A router holds context. A central router layer receives qualifying events, maintains state across the task queue, and assigns work — this is what lets the agent track an ongoing problem instead of treating every message as a cold start.
4. The agent works in its own session. Once assigned, Super-Agent opens an isolated session to investigate and fix the issue, the same separation-of-concerns pattern Claude Code and Codex already use.
5. It replies where the question was asked. Results — commonly a pull request — land back in the original GitHub thread or chat, not in a separate dashboard someone has to remember to check.

The performance numbers DeepSeek is circulating are early and self-reported, but they’re specific: team trials put the parallel, coordinator-based architecture at a median task completion of 12.4 minutes, with a 95th-percentile time of 15.7 minutes — against 29.5 minutes at the 95th percentile for a conventional, one-task-at-a-time setup. Call it roughly twice as fast on the slow end, with the obvious caveat that DeepSeek ran its own benchmark on its own agent.

Why the Timing Matters

The turnaround is the part worth sitting with. Per Bloomberg and SCMP’s reporting on the October 31 drop, what shipped is a working router, an event-filtering layer, and integrations into two separate chat surfaces — GitHub and Feishu — built from scratch in the time the Harness team has existed. That’s a fast build cycle for an agent runtime, even one the company itself calls unfinished.

Anthropic didn’t build Claude Code’s moat on raw capability alone — plenty of agents can write a pull request. The moat is distribution, reliability at scale, and a subscription model enterprises already trust enough to put on a procurement form. Claude Code Routines, which Anthropic shipped back in April, is Anthropic’s own answer to “work that happens without a human in the loop” — scheduled, API-triggered, or GitHub-event-triggered workflows running on Anthropic’s cloud. It’s a mature, supported feature with usage caps tied to your plan. Super-Agent is an unsupported prototype with an MIT license and no usage cap at all, because there’s no company standing behind it the way Anthropic stands behind Claude Code.

That gap is exactly where the pressure lands. A free, self-hosted tool doesn’t need to match Claude Code’s reliability to matter — it just needs to be good enough that some fraction of Claude Code’s user base starts asking why they’re paying for a subscription tier. DeepSeek has run this play before: its V4-Pro API pricing cut made permanent in May put a public, 6-7x cost gap in front of every enterprise buyer evaluating US frontier labs. Super-Agent is the same logic applied to tooling instead of tokens — don’t compete on capability, compete on the fact that the alternative costs nothing.

The Harness Team’s Bigger Bet

DeepSeek assembled its Harness team earlier in 2026 as an independent business line with an explicit mandate: benchmark Claude Code and build a competing agent framework. “Model + Harness = Agent” is reportedly the internal framing — the idea that the underlying language model matters less than the scaffolding of tools, memory, sub-agent orchestration, and termination logic wrapped around it. Super-Agent is the most concrete public evidence yet that the team is executing against that mandate on a real timeline, not just talking about it.

That framing should sound familiar to anyone who’s followed Anthropic’s own multi-agent harness work this year. Both companies have landed on the same conclusion: the model is table stakes, and the orchestration layer around it is where the next round of competitive advantage gets built. The difference is who gets to see the blueprint. Anthropic’s harness innovations ship as features inside a paid product. DeepSeek just put its version of the blueprint on GitHub for anyone to fork.

The Part DeepSeek Isn’t Hiding

Give DeepSeek credit here: the documentation doesn’t pretend proactive autonomy is free of risk. The project’s own materials warn that “an agent that acts proactively needs carefully scoped permissions, audit trails, interruption controls, and clear human approval points,” and flag that malicious instructions injected into a monitored channel could compromise connected tools. That’s not boilerplate — it’s a direct acknowledgment that an agent reading your team’s chat and opening pull requests unsupervised is a bigger attack surface than an agent that waits in a sidebar for you to type something.

This isn’t a hypothetical concern specific to DeepSeek. Claude Code has had its own rogue-agent moments in 2026, and any tool that acts on unvetted input — a chat message, an issue comment — inherits prompt-injection risk by design. Independent research on multi-agent coding setups has also found that parallelized agents can underperform a single careful agent on tasks that require tight coordination, so “faster” and “more autonomous” don’t automatically mean “more correct.” Super-Agent’s own numbers come from DeepSeek’s internal trials, not an independent benchmark, and the company’s repo explicitly says the code “may contain bugs” and isn’t production-ready. Treat the speed claim as a prototype’s self-report until someone outside DeepSeek reproduces it.

Our Take

I don’t think Super-Agent beats Claude Code on raw capability today, and DeepSeek isn’t claiming it does — “research prototype,” right there in the README, is about as honest as a competitive release gets. The question I keep coming back to isn’t pricing. It’s architecture: is monitoring a team’s existing chat and jumping in before anyone asks a genuine capability gap between DeepSeek and Anthropic, or just a demo that looks impressive until it’s wired into a repo that actually matters?

My read: it’s a real idea, not a repackaged one. Claude Code and Codex both default to request-response — open a session, ask, get an answer. Claude Code Routines adds scheduled and event-triggered runs on top of that, but someone still has to configure the trigger. Super-Agent’s router-plus-filter design is a different shape entirely: the agent is already watching, already triaging, and only opens a full session once a human flags something worth fixing in the channel they already use. That’s a meaningfully different interaction model, not just an open-source clone of a feature Anthropic already shipped.

Where I’d push back on DeepSeek’s own framing: proactive monitoring of unvetted chat is exactly where the attack surface grows, and a 12.4-minute median completion time is DeepSeek grading its own homework. Until a team outside DeepSeek reproduces those numbers on a codebase that isn’t DeepSeek’s, I’d file this as a credible proof of concept for a new agent shape — not evidence that the shape is ready to replace what Claude Code or Codex already do well on hard, high-stakes work.

Frequently Asked Questions

What is DeepSeek’s Super-Agent?

Super-Agent (repo name dsh-superagent) is an open-source, MIT-licensed coding agent that DeepSeek’s Harness team released on GitHub on October 31, 2026. Unlike typical request-response coding assistants, it monitors GitHub activity and Feishu team chats for coding problems and proposes fixes — including pull requests — without being explicitly summoned for each task.

Is Super-Agent an official DeepSeek product?

No. DeepSeek’s own repository documentation describes it as a “research prototype for reference only, not an official DeepSeek product,” built largely from the team’s internal day-to-day workflows and explicitly not production-ready.

How is Super-Agent different from Claude Code?

Claude Code is a paid, supported product that operates primarily on explicit requests or scheduled Routines. Super-Agent’s distinguishing feature is proactive monitoring — it reads ongoing team chat for problems and can start work before anyone directly asks it to, through a router layer that filters events and maintains context across a task queue.

How long has DeepSeek been building toward Super-Agent?

DeepSeek staffed its Harness team as an independent business line earlier in 2026 with an explicit mandate to benchmark Claude Code and build a competing agent framework. The October 31 GitHub release, covered by Bloomberg and SCMP on November 1-2, is that team’s first public artifact — a router, event filter, and chat integrations built in roughly ten months.

Does Super-Agent pose security risks?

DeepSeek’s own documentation says yes, in principle — it warns that proactive agents need “carefully scoped permissions, audit trails, interruption controls, and clear human approval points,” and flags that malicious instructions in a monitored channel could compromise connected tools. Any agent that acts on unvetted chat or issue content carries prompt-injection risk by design, a problem that isn’t unique to DeepSeek.

Should enterprise teams switch from Claude Code to Super-Agent?

Not yet, for anything production-critical. It’s an unsupported, self-admitted prototype with no SLA, no enterprise support channel, and benchmark numbers that come from DeepSeek’s own internal trials rather than independent verification. It’s worth evaluating for internal experimentation or non-critical repos, the same way teams have experimented with other open-source coding tools like Aider.

What does this mean for AI coding tool pricing generally?

It extends a pattern DeepSeek already set with its permanent 75% API price cut in May: compete on cost and openness rather than matching US labs feature-for-feature. Expect more pressure on Anthropic, OpenAI, and other coding-assistant vendors to justify subscription pricing against a free, open alternative — even an imperfect one.

---

Last updated: November 2, 2026. Sources: DeepSeek — dsh-superagent-prototype on GitHub · Bloomberg — DeepSeek Drops Enhanced Coding Agent With Aim to Rival Anthropic · South China Morning Post — DeepSeek unveils AI ‘super agent’ in coding race with OpenAI, Anthropic · Help Net Security · SiliconReport · CCTest.

Related reading: Anthropic’s S-1: $120.5B to Musk’s SpaceX for GPUs · Claude Code Routines: What Enterprises Should Know · DeepSeek’s 75% Price Cut: What It Costs You Now · Anthropic’s Multi-Agent System: 4-Hour AI Dev · Claude Code Self-Hosted: The Fix for Rogue Agents? · DeepSeek V4 Review 2026