DeepSeek Super-Agent vs Claude Code: real?
Published 2 November 2026
Partly true. DeepSeek did publish an open-source "Super-Agent" coding agent that can listen to a team chat, but the public code waits to be asked and, by default, only records the work it would have taken on. The agent that reads group chats and fixes things unprompted is DeepSeek's internal version, and we found no test of the public prototype against Claude Code. If you lead an engineering team, nothing here should change your tooling or hiring plans yet.
What happened
On 31 October 2026 DeepSeek published dsh-superagent-prototype on GitHub: one commit, MIT licence, TypeScript. By 2 November it had about 11,200 stars. The README calls it "a proactive coding agent for one GitHub repository" that "listens to GitHub and, optionally, Feishu". It runs inside DeepSeek Harness, the team's own agent host.
On 1 November DeepSeek Harness linked its design essay on X, "Exploring Super-agent Design". A search preview shows the essay calling the plugin "only a prototype for validating the technical approach, not an official DeepSeek product".
Bloomberg framed it as a challenge to Anthropic's Claude Code that can "suggest solutions without being summoned". We saw that line only in a search preview, as the article is paywalled. A Cointelegraph post turned it into an agent that "monitors developer group chats and suggests solutions without prompting".
Is it true?
We read the docs, then on 2 November cloned the repository at its only commit, 6337fc8, and read src/policy.ts and src/sources/feishu.ts. The defaults match the docs.
| Claim in the coverage | What the public code does |
|---|---|
| Monitors group chats | Feishu only, and only chats you list. Other group chats are ignored unless you set them to "mentions". A listed chat reads every message only if set to "all". |
| Acts without prompting | Default mode is "observe": unrequested work is assessed and logged as "would accept", and nothing starts. |
| Starts work on its own | Only in "auto" mode, and only when an enabled rule matches. A few rules ship enabled, so check them first. |
| Rivals Claude Code | Bloomberg's framing. We found no head-to-head test. |
The philosophy doc says: "It acts only when asked, or when a written rule allows it." The README also warns that the code "is not identical to our internal version" and that some designs in the essay "are not implemented in this prototype".
The unprompted behaviour comes from the essay's account of DeepSeek's internal instance, which, according to a TechBytes summary, "monitored public group chats without direct mentions" and opened pull requests. We could not open the essay, so that is second-hand.
The same applies to the number people quote: a drop in 95th-percentile task time from 29.5 to 15.7 minutes, described on X as DeepSeek's "own test". We could not read the method, so we cannot reproduce it.
Is it useful for engineering teams?
Not as a product yet. DeepSeek calls it "research code, not production-ready software". The code ships connectors for GitHub and Feishu only, so Slack or Microsoft Teams users get no chat side.
Its permission model is worth reading. It has an allow-list of people who can ask, a log of what the agent would have done, and written rules before any unrequested work. Each request runs "in its own agent session and directory".
Anthropic describes Claude Code as a tool that "reads your codebase, edits files, runs commands", and the same page covers automated review in CI through GitHub Actions or GitLab and a Slack integration. We found no test that compares the two.
What do the critics say?
The sharpest reaction is from @teortaxesTex: "As written, SuperAgent is something of a hot mess." The post argues that DeepSeek "evolved internal crutches" for it and did not try to make it generally robust. @SkylerMiao7 called it "neither new nor very usable yet". Both are opinions without a test.
The TNW report adds a caution from the CooperBench study: coding agents working together scored "on average 30% lower than when each agent did the tasks alone", though the study did not test DeepSeek's agent. By 2 November we found no independent hands-on test.
For a buyer, prompt injection is the bigger open question. The docs say chat messages "cannot change those permissions", but we found no public test of that claim, or of what the agent could expose from a chat it reads.
What it means if you are hiring engineers
Any agent that reads team chat sends its pull requests to your reviewers. In a technical interview, give candidates an agent-written pull request with one subtle bug planted and ask what they would merge.
Before you wire an agent into team chat, check what people paste there, because the agent will read it. Someone also has to write the allow-lists and rules and run them in observe mode first. That is work for platform and DevOps engineers or an AI team, not for whoever installs the tool. Wild.Codes is an AI hiring platform for engineering teams, with no cost until you hire.
FAQ
Does DeepSeek Super-Agent read my team chats and jump in on its own?
Not by default. The public prototype reads GitHub, direct messages to its Feishu bot and only the Feishu group chats you list. In default observe mode it only logs unrequested work.
Is DeepSeek Super-Agent free to self-host?
The code is MIT-licensed. You still need DeepSeek Harness, a model provider and several developer tools.
Is it a Claude Code replacement?
No test shows that. We found no comparison of the two, and DeepSeek calls its public version research code.
How we checked
Wild.Codes wrote this. We help companies hire engineers. We pay for Anthropic's Claude models in our own work and have no other commercial relationship with Anthropic, and none with DeepSeek. On 2 November 2026 we read the repository's README, design docs and setup guide, cloned commit 6337fc8 without running it and checked the default settings in the source. Bloomberg's article and DeepSeek's essay were closed to us.
Sources
- deepseek-ai, dsh-superagent-prototype (repository and README), GitHub, https://github.com/deepseek-ai/dsh-superagent-prototype, created 31 Oct 2026, read 2 Nov 2026
- deepseek-ai, README.md, GitHub, https://raw.githubusercontent.com/deepseek-ai/dsh-superagent-prototype/main/README.md, 31 Oct 2026, read 2 Nov 2026
- deepseek-ai, docs/000-philosophy.md, GitHub, https://raw.githubusercontent.com/deepseek-ai/dsh-superagent-prototype/main/docs/000-philosophy.md, 31 Oct 2026, read 2 Nov 2026
- deepseek-ai, docs/002-authorization.md, GitHub, https://raw.githubusercontent.com/deepseek-ai/dsh-superagent-prototype/main/docs/002-authorization.md, 31 Oct 2026, read 2 Nov 2026
- deepseek-ai, SETUP.md, GitHub, https://raw.githubusercontent.com/deepseek-ai/dsh-superagent-prototype/main/SETUP.md, 31 Oct 2026, read 2 Nov 2026
- DeepSeek Harness, post linking "Exploring Super-agent Design", X, https://x.com/DeepSeekHarness/status/2116857257245987236, 1 Nov 2026, read 2 Nov 2026 (essay quote from a search preview)
- Bloomberg, "DeepSeek Drops Enhanced Coding Agent With Aim to Rival Anthropic", https://www.bloomberg.com/news/articles/2026-11-01/deepseek-drops-enhanced-coding-agent-with-aim-to-rival-anthropic, 1 Nov 2026, search preview read 2 Nov 2026
- Bloomberg (@business), post on X, https://x.com/business/status/2116754642222616835, 1 Nov 2026, read 2 Nov 2026
- Cointelegraph, post on X, https://x.com/Cointelegraph/status/2116885723601068339, 1 Nov 2026, read 2 Nov 2026
- TNW, DeepSeek Super-Agent report, https://thenextweb.com/news/deepseek-super-agent-open-source-coding-agent, 1 Nov 2026, read 2 Nov 2026
- TNW, CooperBench study (arXiv 2601.13295) as reported in the same article, https://thenextweb.com/news/deepseek-super-agent-open-source-coding-agent, 1 Nov 2026, read 2 Nov 2026
- TechBytes, "Exploring Super-agent Design: Towards a More Autonomous Agent System" (summary), https://techbytes.app/posts/exploring-super-agent-design-towards-a-more-autonomous-agent-system/, 1 Nov 2026, read 2 Nov 2026
- @kimmonismus, post on X, https://x.com/kimmonismus/status/2116905923109089285, 1 Nov 2026, read 2 Nov 2026
- @teortaxesTex, post on X, https://x.com/teortaxesTex/status/2116988654602031170, 1 Nov 2026, read 2 Nov 2026
- @SkylerMiao7, post on X, https://x.com/SkylerMiao7/status/2117118200819404901, 2 Nov 2026, read 2 Nov 2026
- Anthropic, Claude Code overview, https://code.claude.com/docs/en/overview, updated 20 Oct 2026, read 2 Nov 2026
• PHP expertise;
• Database management skills;
•Jungling traits, methods, objects, and classes;
• Agile & Waterfall understanding and use;
• Soft skills (a good team player, high-level communication, excellent problem-solving background, and many more)
• OOP & MVS deep understanding;
• Knowledge of the mechanism of how to manage project frameworks;
• Understanding of the business logic the project meets;
• Cloud computing & APIs expertise.
• Reasonable life-work balance;
• The opportunity to implement the server-side logic via Laravel algorithms;
• Hassle-free interaction with back-end and front-end devs;
• Strong debugging profile.
• Using HTML, XHTML, SGML, and similar markup languages
• Improving the usability of the digital product
• Prototyping & collaboration with back-end JS experts
• Delivery of high-standard graphics and graphic-related solutions
• Using JS frameworks (AngularJS, VueJS, ReactJS, etc
• Clean coding delivery and timely debugging & troubleshooting solution delivery
• UI testing and collaboration with front-end JS teammates
• Database experience
• Building APIs while using REST or similar tech solutions
• Collaboration with project managers and other devs
• Delivery of design architecture solutions
• Creation of designs & databases
• Implementation of data protection and web cybersecurity strategies.
• Both front-end and back-end qualifications
