What this is
DeepSeek has open-sourced Harness (dsh), an Agent orchestration framework built around the core design principle of "everything is a plugin." Our read on what this means: competition among China's large model companies has moved from the model layer down to the toolchain layer.
Specifically, user-submitted requests are received by a "scheduling hub," broken down, distributed to different plugins for execution, and finally aggregated and returned. The architecture references the paper "A Programming Paradigm for Spatiotemporal Composability" and runs on Cordis at the base layer. The framework offers three entry points: a Web interface (pnpm dsh web), a command line, and a Python SDK (for automation scripts).
Official docs explicitly state that developing plugins should "leverage AI programming tools to complete," effectively letting AI write the toolchain for AI. One easy-to-miss detail: DeepSeek's API now uses peak/off-peak pricing, hinting that long-running Agent tasks are most cost-effective when run during nighttime off-peak hours.
Industry view
We note that this is the first time a Chinese large model company has seriously built tooling at the "Agent operating system" tier. Prior competition centered on the model itself—who scores higher on benchmarks, who prices lower. The pivot to toolchains follows the same logic as Anthropic's MCP (a standard protocol letting models call external tools) and OpenAI's Agents SDK: the model is the engine, but the toolchain is the transmission.
However, plenty of voices urge caution. First, "everything is a plugin" sounds elegant, but the cost is that every plugin must be written by the user—developer-friendly, but enterprise-IT-unfriendly. Second, the Agent orchestration track is already crowded; LangChain, AutoGen, and CrewAI are more mature, and DeepSeek won't find it easy to capture developer mindshare. Third, peak/off-peak pricing is clever, but enterprise long-running tasks often need to respond to real-time business, so the cost savings are limited.
Impact on regular people
For enterprise IT: no rush in the short term. This framework is currently more "a toy for developers" than "a product for enterprises." Worth watching, not worth allocating resources immediately.
For individual professionals: if you can code, this is a low-cost entry point for experimenting with Agent development. The Python SDK, combined with DeepSeek's relatively cheap API, lets you build a small assistant without burning cash.
For the consumer market: no visible impact yet. The winner of the Agent toolchain race hasn't emerged—wait for a few benchmark cases before judging.