DeepSeek Open Sources the Architecture Behind Its Super-Agent

DeepSeek published deepseek-harness, an MIT-licensed GitHub repository containing the full architecture of its always-on autonomous "Super-agent," on August 13, 2026, and the repo reached 177,904 stars within eight days and passed 188,000 by late October, per tracking cited by levelup.gitconnected.com. The harness runs on a micro-kernel plugin runtime called Cordis, installs via `npx @deepseek-ai/dsh web` with a local UI at 127.0.0.1:3080, and lets a primary agent delegate code review, resource search and testing to sub-agents while logging every model call, tool invocation and context-compression step for replay or audit. DeepSeek's own documentation warns the preview is iterating rapidly with compatibility-breaking changes ahead, and the Super-agent prototype ships in a separate repository, dsh-superagent-prototype, built on top of the runtime.

DeepSeek isn't selling access to an autonomous agent. It's handing out the blueprint for free, and asking developers to build their own. On August 13, 2026, DeepSeek pushed a new repository to GitHub called deepseek-harness. Within eight days it had racked up 177,904 stars, according to tracking cited by levelup.gitconnected.com, and by late October that count had climbed past 188,000, with some trackers citing figures north of 200,000. The repository holds the full architecture behind what DeepSeek calls Super-agent, an always-on autonomous agent, released as a working prototype under an MIT license. That license matters. Claude Code and OpenAI's Codex are agent products you rent through an API key and a subscription. DeepSeek Harness, often shortened to dsh, is something you can clone, inspect, and rebuild from the inside. Install Node.js and run npx @deepseek-ai/dsh web . A local web UI opens at 127.0.0.1:3080. From there every piece of the system, the model adapters, the tool registry, session logging, sandboxing, scheduling, even the interface itself, is a swappable plugin. The plugin runtime underneath Harness is called Cordis, described in DeepSeek's own documentation as "A Programming Paradigm for Spatiotemporal Composability." In plain terms, it's a micro-kernel that lets an agent keep running, keep logging, and keep delegating without a human kicking off each new loop. According to reporting from The New Stack, a primary agent inside Harness can hand off subtasks like code review, resource search, and testing to specialized sub-agents. It then aggregates their results and adjusts strategy on its own. That's a meaningful step past the AutoGPT-style agents that dominated headlines a few years back. Those ran a single loop: think, act, observe, repeat, until a goal was met or a budget ran out. Anthropic's computer use and OpenAI's Operator extended that into clicking and typing on a screen, but both still start and stop around a task a human defines. DeepSeek is after something else. What it's publishing is closer to infrastructure for persistence itself: every model call, tool invocation, and context-compression step gets logged so it can be replayed or audited later, which is the kind of observability enterprise buyers actually ask for before they'll trust an agent left running unattended. China's open AI models now trail the US frontier by just seven months https://startupfortune.com/chinas-open-ai-models-now-trail-the-us-frontier-by-just-seven-months/ Epoch AI's Capabilities Index, built from 39 benchmarks including FrontierMath and SWE-Bench Verified, shows DeepSeek, Qwen and Kimi closing a gap that stood at 17.5 points on MMLU in 2023. - how China's AI models catch up to US https://startupfortune.com/chinas-open-ai-models-now-trail-the-us-frontier-by-just-seven-months/ - China open source AI models trailing US frontier https://startupfortune.com/chinas-open-ai-models-now-trail-the-us-frontier-by-just-seven-months/ Composio's technical comparison frames the contrast bluntly: where Claude Code and Codex hand you a finished product, DeepSeek Harness hands you a frame and expects you to build the walls. It's rougher out of the box. There's no polish, no onboarding flow, and DeepSeek says in its own docs that the preview is iterating rapidly with compatibility-breaking changes still ahead. A separate repo for the agent itself The Super-agent prototype lives in its own repository, dsh-superagent-prototype, built on top of the Harness runtime rather than folded into it. That separation is itself a signal. DeepSeek isn't shipping one finished autonomous agent and calling it done. It's shipping the plumbing, then showing one worked example of what you can build with it, and inviting the ecosystem to build the next ones. That's a direct shot at the business model Western labs are currently selling. Anthropic and OpenAI both charge a premium for packaged agent capability on top of their models, Opus 4.7 running $5 per million input tokens and $25 per million output, against GPT-5.5 at $5 and $30, figures reported by CloudZero's pricing breakdown. DeepSeek's own V4-Pro model has moved in the opposite direction, dropping as low as $0.0035 per million tokens on a 75% permanent cut announced earlier this year, according to Engadget's coverage. The irony is that DeepSeek later raised V4-Pro's API prices again, a reversal flagged by TheNextWeb, even as it kept giving away the agent architecture itself for nothing. The pricing lever and the open-source lever are now moving independently, and that's worth watching: DeepSeek doesn't need to win on API price if it can win by becoming the substrate every other agent framework gets built on top of. For the agent-framework startups that have spent the past two years building orchestration layers on top of GPT and Claude, a free, MIT-licensed, enterprise-grade reference architecture from a well-funded competitor is a genuine threat, not a curiosity. Anthropic has previously accused DeepSeek of running "distillation attacks" against its models. Giving away the architecture for persistent agents, for free, under a permissive license, with nearly 200,000 developers already watching the repo, is the kind of move that forces a response rather than waits for one. Also read: Judge Lets NAACP's Pollution Lawsuit Against Elon Musk's xAI Proceed https://startupfortune.com/judge-lets-naacps-pollution-lawsuit-against-elon-musks-xai-proceed/ • Xbox Manager Says Studio Uses AI Widely Then Deletes His Account https://startupfortune.com/xbox-manager-says-studio-uses-ai-widely-then-deletes-his-account/ • Comcast cut 30 million service calls with AI as it preps cable layoffs https://startupfortune.com/comcast-cut-30-million-service-calls-with-ai-as-it-preps-cable-layoffs/ This article is posted in AI News https://startupfortune.com/category/ai/ , check it out for more related stories. Join the discussion Open in the community → https://startupfortune.com/community/ Almost there. Sign in and your reply posts straight away.