--- title: DeepSeek open-sources Super-Agent, an experimental... description: DeepSeek open-sourced Super-Agent, an experimental coding agent that looks aimed at Claude Code. The code is on GitHub as dsh-superagent under an MIT license.... canonical: https://daily.dev/posts/deepseek-open-sources-super-agent-an-experimental-coding-agent-that-listens-to-team-chats-q8puk0oia twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: DeepSeek open-sources Super-Agent, an experimental coding agent that listens to team chats | daily.dev og:description: DeepSeek open-sourced Super-Agent, an experimental coding agent that looks aimed at Claude Code. The code is on GitHub as dsh-superagent under an MIT license.... og:url: https://daily.dev/posts/deepseek-open-sources-super-agent-an-experimental-coding-agent-that-listens-to-team-chats-q8puk0oia og:image: https://api.daily.dev/og/posts/Q8pUk0OIa.png og:image:alt: DeepSeek open-sources Super-Agent, an experimental coding agent that listens to team chats og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # DeepSeek open-sources Super-Agent, an experimental coding agent that listens to team chats **[Collections](https://daily.dev/sources/collections)** · 2 min read · 5 upvotes · 0 comments ## Summary DeepSeek open-sourced Super-Agent, an experimental coding agent that looks aimed at Claude Code. The code is on GitHub as dsh-superagent under an MIT license. It works on a single GitHub repository and can listen to Feishu chats, then answer tasks in the original thread, for example by opening a pull request. DeepSeek calls it a research prototype, not an official product, with the goal of taking as little human attention as possible while giving as much value as possible. DeepSeek also published a design write-up on exploring a more autonomous agent system, which @omarsar0 called 'an absolute masterclass'. A separate CooperBench study found agents working as a team had success rates 30% lower than the same agents working alone, though it didn't test Super-Agent. ## Content DeepSeek has open-sourced Super-Agent, an experimental coding agent that looks aimed at Anthropic's Claude Code. The code is on GitHub as dsh-superagent under an MIT license. ## What it does Super-Agent works on a single GitHub repository. It can listen to Feishu chats and answer tasks in the original thread, for example by opening a pull request. DeepSeek calls it a research prototype, not an official product. Its stated goal is to "take as little human attention as possible, and give as much value as possible." ## The design write-up DeepSeek also published an article, "Exploring Super-agent Design: Towards a More Autonomous Agent System." @Hesamation put together a visual explainer of it. @omarsar0 called it "an absolute masterclass" in building a modern harness, and said it is the best description he has read of how he already works with agent teams on his own custom harness. The argument, as @omarsar0 summarizes it: - The agent should be able to run continuously and stay productive across tasks. That calls for a coordination-execution architecture, which is now realistic because LLM agents are better at coordination and at keeping persistent context. - Once the core architecture works, you can add proactive pieces such as event triggering and acting ahead of time, plus new ways to talk to agents, like an asynchronous ReAct loop. - Scaling depends on state management and sandboxes. The design should separate state, compute, and environment, and it should persist and stay available across devices. - None of this is tied to DeepSeek's harness. Other existing harnesses can do it too. - As agents get better, people should look for new ways to interact with them and avoid getting in their way. ## A caution on teams of agents One data point cuts against the multi-agent pitch. A CooperBench study found that coding agents working as a team had success rates 30% lower than the same agents working alone. The study did not test Super-Agent, so it says nothing direct about this release. It does show the coordination problem DeepSeek's design is trying to solve is a real one. ## Questions this post answers ### What is DeepSeek's Super-Agent and how is it different from Claude Code? Super-Agent is an experimental, open-source coding agent from DeepSeek, released on GitHub as dsh-superagent under an MIT license. It works on a single GitHub repository and can listen to Feishu chat threads and act on tasks there, including opening pull requests. DeepSeek describes it as a research prototype rather than an official product, positioned against Anthropic's Claude Code. _Developers comparing coding agent harnesses can track new open-source entrants like this one on daily.dev._ ### Do multi-agent coding setups actually perform better than a single agent working alone? No, a CooperBench study found that coding agents working as a team had success rates 30% lower than the same agents working alone. This finding wasn't tested against DeepSeek's Super-Agent specifically, but it highlights a real coordination problem that agent-team designs need to solve before multi-agent coding setups can be trusted over single-agent approaches. _Anyone deciding between single-agent and multi-agent coding setups can follow benchmark findings like this on daily.dev._ ### What architecture does DeepSeek propose for building more autonomous coding agents? DeepSeek proposes a coordination-execution architecture that lets an agent run continuously and stay productive across tasks, enabled by LLM agents getting better at coordination and persistent context. The design adds proactive elements like event triggering and acting ahead of time, an asynchronous ReAct loop for communication, and separates state, compute, and environment so the system can scale and persist across devices. _Engineers designing their own agent harness can follow architecture write-ups like this on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 2 discussions and 106 comments across x (as of 2026-11-02). **TL;DR:** Developers find the router/worker architecture and separation of state, compute, and environment genuinely interesting, but much of the discussion is Twitter-style low-signal chatter, with the more substantive replies focused on open design questions like conflict resolution, handoffs, and token costs rather than clear praise or criticism. **Sentiment:** 30% positive · 45% mixed · 25% skeptical **The case for** - The router dispatching to isolated workers while using a cheap model for routing and a strong model only where needed is seen as a smart design split. - Separating state, compute, and environment so a crashed worker can resume instead of restarting is viewed as a key architectural win. - Letting agents run on their own schedule and only pulling in a human for decisions is seen as reducing operational overhead. - The chain-of-recall / lazy-loading context approach is appreciated as a practical way to avoid context window bloat. **The pushback** - Several people worry about how conflicting results from parallel workers get reconciled without just picking the loudest answer. - An agent acting proactively without being asked is flagged as risky, especially for anything customer-facing, without clear sign-off/approval gates. - The setup is noted to consume very large numbers of cache-read tokens per run, raising concerns about API cost versus time saved. - Multiple commenters note persistent state can drift or go stale, and it's unclear how the system versions or expires it. - Some question whether a sub-agent's claim of being 'done' is actually verified rather than just trusted. **By community** - x (mixed): A flood of mostly low-substance replies mixed with a smaller set of thoughtful architecture questions about routing, state separation, and conflict handling, landing on cautious interest rather than a clear verdict. **Hottest debate:** Whether the router/worker split with separated state truly solves coordination and conflict problems, or just moves the hard problems (verification, conflicting edits, stale state) somewhere less visible. **Open questions** - How does the router decide which prior turns or context to discard versus re-run for each worker? - How are conflicting edits or results between parallel workers reconciled rather than just trusting the loudest one? - How does the system verify a sub-agent's work or completion claim rather than taking its word for it? - How is persistent state versioned or expired so agents don't keep acting on stale assumptions? - Where exactly should the line be drawn between autonomous action and requiring human approval, especially for external-facing actions? **Highlights** > @Hesamation The visual is clean, but the real cost hides in the loop: every tool call is a network round-trip, and each result re-enters the context window. At 20+ steps, latency and token spend compound faster than the diagram suggests. > — [fishjar on x](https://x.com/fishjar/status/2116980679082983847) > @omarsar0 he Super-agent concept works brilliantly, but at the cost of massive token consumption (23.2M cache-read tokens per run). Saving developer time is great, but maintaining isolated workers and constant caching inflates API bills fast > — [od\_chenko on x](https://x.com/od_chenko/status/2116960684533862683) > @omarsar0 A coordination-execution split only holds if state lives outside the context window. Otherwise long-running agents just re-read their own history every turn, and prefill cost plus context rot eat the autonomy gains. > — [flaviusburca on x](https://x.com/flaviusburca/status/2116955492652380346) > @Hesamation The "acts without being asked" box is where I'd slow down. One agent across chat, email and code is great until it decides on its own to reply to a customer. I let it act alone on internal work and draft for approval on anything external. Curious how they draw that line. > — [jeremyblocknova on x](https://x.com/jeremyblocknova/status/2117012659992498518) > @Hesamation the context split is the part i'd want to dig into. if each worker only sees its slice of the repo, how does the lead catch two workers making conflicting edits? faster completion is nice but that's usually where these setups break > — [Dakshagarwal76 on x](https://x.com/Dakshagarwal76/status/2116939275769884788) **Source threads** - [x](https://x.com/Hesamation/status/2116920980002230772) · 0 points · 52 comments - [x](https://x.com/omarsar0/status/2116942815150157871) · 0 points · 54 comments ## Similar posts on daily.dev - [GitHub - aattaran/deepclaude: Use Claude Code's autonomous agent loop with DeepSeek V4 Pro, OpenRouter, or any Anthropic-compatible backend. Same UX, 17x cheaper.](https://daily.dev/posts/github---aattaran-deepclaude-use-claude-code-s-autonomous-agent-loop-with-deepseek-v4-pro-openrout-0rcoomwtj) · Hacker News · 3 upvotes · 0 comments - [DeepSeek is hiring 150 engineers, and none of them will touch a model](https://daily.dev/posts/deepseek-is-hiring-150-engineers-and-none-of-them-will-touch-a-model-lv5ahjhsh) · The New Stack · 0 upvotes · 0 comments --- Tags: [#open-source](https://daily.dev/tags/open-source), [#ai-agents](https://daily.dev/tags/ai-agents), [#claude-code](https://daily.dev/tags/claude-code), [#deepseek](https://daily.dev/tags/deepseek) [View this post on daily.dev](https://daily.dev/posts/deepseek-open-sources-super-agent-an-experimental-coding-agent-that-listens-to-team-chats-q8puk0oia) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"DeepSeek open-sources Super-Agent, an experimental coding agent that listens to team chats","url":"https://daily.dev/posts/deepseek-open-sources-super-agent-an-experimental-coding-agent-that-listens-to-team-chats-q8puk0oia","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/deepseek-open-sources-super-agent-an-experimental-coding-agent-that-listens-to-team-chats-q8puk0oia"},"datePublished":"2026-11-01T17:18:57.605Z","dateModified":"2026-11-02T00:50:09.797Z","description":"DeepSeek open-sourced Super-Agent, an experimental coding agent that looks aimed at Claude Code. The code is on GitHub as dsh-superagent under an MIT license....","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/deepseek-open-sources-super-agent-an-experimental-coding-agent-that-listens-to-team-chats-q8puk0oia","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":5},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"open-source,ai-agents,claude-code,deepseek","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"DeepSeek open-sources Super-Agent, an experimental coding agent that listens to team chats"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/deepseek-open-sources-super-agent-an-experimental-coding-agent-that-listens-to-team-chats-q8puk0oia#faq","mainEntity":[{"@type":"Question","name":"What is DeepSeek's Super-Agent and how is it different from Claude Code?","acceptedAnswer":{"@type":"Answer","text":"Super-Agent is an experimental, open-source coding agent from DeepSeek, released on GitHub as dsh-superagent under an MIT license. It works on a single GitHub repository and can listen to Feishu chat threads and act on tasks there, including opening pull requests. DeepSeek describes it as a research prototype rather than an official product, positioned against Anthropic's Claude Code. Developers comparing coding agent harnesses can track new open-source entrants like this one on daily.dev."}},{"@type":"Question","name":"Do multi-agent coding setups actually perform better than a single agent working alone?","acceptedAnswer":{"@type":"Answer","text":"No, a CooperBench study found that coding agents working as a team had success rates 30% lower than the same agents working alone. This finding wasn't tested against DeepSeek's Super-Agent specifically, but it highlights a real coordination problem that agent-team designs need to solve before multi-agent coding setups can be trusted over single-agent approaches. Anyone deciding between single-agent and multi-agent coding setups can follow benchmark findings like this on daily.dev."}},{"@type":"Question","name":"What architecture does DeepSeek propose for building more autonomous coding agents?","acceptedAnswer":{"@type":"Answer","text":"DeepSeek proposes a coordination-execution architecture that lets an agent run continuously and stay productive across tasks, enabled by LLM agents getting better at coordination and persistent context. The design adds proactive elements like event triggering and acting ahead of time, an asynchronous ReAct loop for communication, and separates state, compute, and environment so the system can scale and persist across devices. Engineers designing their own agent harness can follow architecture write-ups like this on daily.dev."}}]} ```