DeepSeek Harness developer preview

Most of the interest landed on that traceability. People read it as a practical break from the black-box style of OpenAI and Anthropic agent products, where hidden reasoning and partial logs make it harder to debug workflows, tune prompts, or understand why a tool chain failed. Several readers also pointed out that the feature itself is not novel. It is structured event logging or event sourcing applied to agent runs. What felt new was seeing a major model vendor expose it directly in a first-party harness instead of hiding it behind product constraints. The Cordis side got a more mixed reaction. People with experience in plugin systems said the core ideas are familiar: explicit init and cleanup hooks, dependency graphs, and unloadable modules look a lot like OSGi, Eclipse plugins, React effect cleanup, or other older lifecycle systems. The favorable read was that DeepSeek packaged those ideas for agent tooling, where in-session tool generation, hot reload, and rapid experimentation matter more than in ordinary apps. The skeptical read was that this adds a lot of machinery for cases where most plugins are simple and mostly independent. That split carried into the broader verdict on “everything is a plugin.” Supporters liked the flexibility, especially if agents can generate custom plugins on demand and if DeepSeek’s own models were post-trained against this harness style. Critics saw the usual plugin-platform risks right away: weak defaults, ecosystem churn, dependency breakage, and users spending more time wiring extensions than getting work done. A lot of readers also thought the packaging story undercut the pitch, pointing to a thin README, confusing docs, Node and npm baggage, and very large install size for what is still a rough v0.1. The result was interest, not awe. People came away thinking DeepSeek shipped a serious agent substrate with strong observability and some credible systems ideas, but not a clearly superior end-user coding product yet.