On August 13, DeepSeek did something unusual: instead of open-sourcing a model, it open-sourced the scaffolding around the model—the Agent orchestration layer Harness plus the protocol layer Cordis, free under the MIT license (the most permissive open-source license). This is the first time a Chinese large-model company has bet on AI infrastructure rather than the model itself.
What this is
Harness turns DeepSeek models into the orchestration layer for production-ready Agents (AI assistants that autonomously execute tasks), accompanied by a paper and the meta-framework Cordis. The core shift is "protocol-first" replacing "framework-first." LangChain's agent scheduling loop is privileged code—you can only fork it (copy the source and maintain it independently) to make changes. Cordis turns every component into an ordinary plugin that follows the same protocol. Want to swap scheduling logic? Write a new plugin registering the same service name, and the old one is automatically replaced.
The baseplate is the protocol; the blocks are plugins. Koishi, the verification case DeepSeek cites, has accumulated over 4,000 community plugins using the same protocol—no walled garden.
Industry view
We note supporters are calling this the "USB-C moment for Agent infrastructure." Developers are no longer held hostage by a single framework, and cross-project plugin reuse becomes possible. The MIT license plus open-source paper lets commercial companies embed it into products without worry.
But the skepticism is equally sharp. First, a pretty protocol doesn't guarantee an ecosystem. LangChain didn't monopolize because of a bad protocol—its dominance comes from first-mover advantage, comprehensive documentation, and locked-in enterprise customers. For Cordis to break through, leading SaaS (software-as-a-service) vendors would need to write their integrations as Cordis plugins—and they have no incentive to leave established systems.
The second risk is more fundamental: Agent frameworks are heavily commoditized, and choosing LangChain versus Cordis doesn't matter much to most enterprises. What actually decides winners is model capability and pricing. This move's political significance outweighs its commercial value—DeepSeek is signaling to the world: "We are not just a model company, we are an infrastructure company."
Impact on regular people
For enterprise IT: selection criteria will change. Previously, evaluating Agent frameworks meant checking feature lists. Going forward, the question is "how big is the plugin marketplace, and how open is the protocol at every layer."
For individual careers: knowing LangChain remains a hot skill, but "understanding the Cordis protocol + being able to write plugins" will be a premium direction in the next 12–18 months.
For consumer markets: end users will notice nothing in the short term. The real beneficiaries are mid-sized companies building AI customer service and AI ops—they can expect customized solutions at lower engineering costs.