Cordis's core code is just six files, tackling the hardest problem in agent engineering: how to make every capability snap together like Lego bricks. That's the signal we see DeepSeek sending by open-sourcing the core module of its Agent framework DSH.
What this is
DSH (DeepSeek Harness) is DeepSeek's internal Agent runtime framework, and Cordis is its underlying dependency. The core design behind this release is called "Everything is Plugin"—turning LLM calls, tools, file systems, Shell, and MCP (a standard protocol that lets AI invoke external tools) into pluggable modules, alongside every other capability.
Traditional Agent frameworks cram everything into a single runtime—change one thing and the whole thing breaks. With a plugin architecture, the core code only handles scheduling, while specific capabilities can be swapped at any time. Each plugin has its own "runtime instance" (Cordis calls it a Fiber), the "capabilities" it exposes (called Service), and a "dependency declaration" (called inject).
One-line summary: a plugin is a module, a Service is what a module exposes to the outside, and inject is how a module tells the system "here's what I need to run." This design addresses the Agent scalability anxiety—scenarios change too fast, and hardcoded code won't survive the next quarter.
Industry view
Supporters argue this aligns with OpenAI's push for the MCP protocol—we believe Agent engineering has entered deep waters, where standardization at the tooling and protocol layer is more worth competing on than the model itself. Chinese AI companies are no longer just racing on parameters; they're now competing on engineering infrastructure—which is good for the developer ecosystem.
But the counterargument is clear. First, "everything is a plugin" isn't new—VS Code and Eclipse proved this model decades ago. Agent plugins, however, must handle permissions, resource cleanup, and cross-plugin communication—complexity an order of magnitude higher than an IDE. Elegant in theory doesn't mean stable in production.
Second, DSH is still too early. Cordis's codebase is tiny, and enterprise-grade stability is unproven. What actually decides whether a framework survives is community ecosystem, not design philosophy—OpenAI's MCP took off because of ecosystem lock-in, not because the protocol itself is elegant.
Third, open source is a double-edged sword. DeepSeek wants to replicate Meta's Llama playbook (open source for ecosystem), but agent frameworks have far shallower moats than models—a competitor can fork and adapt in one to two weeks. What decides the winner is who gets more developers to write plugins on their framework, not who has the prettier design.
Impact on regular people
For enterprise IT: There's now another option for technical evaluation, but we wouldn't recommend going straight to production in the short term. The operational costs of open-source agent frameworks are widely underestimated—the more plugins, the harder version management becomes.
For individual careers: No direct impact yet. But this will accelerate the rollout of AI agents inside enterprises. Over the next year, "knowing how to use Agent frameworks to build workflows" may become a plus for certain roles.
For consumer markets: Ordinary users won't notice directly. But once this infrastructure matures, it will indirectly push up the stability and capability ceiling of more AI products—customer service bots, smart assistants, automation tools.