What this is
On the late night of August 13, DeepSeek quietly open-sourced a project called DeepSeek Harness. This isn't another coding assistant — it's the release, under an MIT license, of the "operating system layer" that AI needs to actually get work done.
There's an industry formula: Agent = Model + Harness. The model thinks; the Harness (literally "tack" — the reins and carriage fitted to a horse) connects the model to the real world — file systems, terminals, the web, APIs, permissions, task planning. These components were previously sealed inside proprietary shells. DeepSeek has open-sourced both the blueprints and the prefab parts, letting developers "tear down any wall and swap any part."
The core architecture is called "everything is a plugin": the kernel, Cordis (a plugin framework from the Koishi.js ecosystem), only handles plugin loading and dependencies. Every concrete capability — model invocation, tools, skills, sessions, Agent loops — is a plugin. Coding is just one preset combination; you can also assemble Agents that have nothing to do with programming.
Industry view
Two signals deserve our attention. First, in the same week DeepSeek switched V4 API pricing to time-of-use peak/off-peak rates: cache-hit prices climbed from roughly ¥0.025 per 1k tokens to about ¥0.30 at peak. With the price tag harder to defend, the company must upgrade from "selling models" to "selling models + execution layer + ecosystem." Second, while 2025 was still about which model was smarter, 2026 competition is visibly tilting toward the Harness — the same model wrapped in different Harnesses produces wildly different experiences depending on what tools it can call and whether it can keep working over time.
But there are cautious voices too. Cordis comes from the relatively niche Koishi.js frontend ecosystem, leaving long-term maintenance and talent pool as open questions. The "everything is a plugin" approach carries a steep learning curve and isn't friendly to non-full-stack developers. Open source doesn't equal free commercialization — thickening the Harness means splitting R&D effort to maintain infrastructure.
Impact on regular people
For enterprise IT: the selection logic has to change. Evaluating AI is no longer just about how smart the model is — it's also about whether the Harness can slot into existing systems and leverage current tools and permissions.
For individual careers: the boundary of collaborating with AI is thickening. The old model was "ask a question, get an answer." The future may be "install a Harness and let AI sit permanently in your workflow." People who understand this layer will have more bargaining power than those who only know how to write prompts.
For consumer markets: end users will notice little in the short term. But within six months to a year, the customer service and office assistants you use will very likely be running on a similar architecture.