DeepSeek Harness makes every agent capability a swappable plugin

DeepSeek Harness is closer to a foundation for assembling agents than a single finished coding agent. Models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the UI are all implemented as plugins. A Cordis kernel manages mounting, unmounting, and dependencies, allowing developers to select or replace capabilities through configuration without changing the Harness source. The developer preview reached 710 points and more than 290 comments on Hacker News, where the clear “Everything is a plugin” design choice gave the discussion a concrete architectural question.

The recording model is equally important. Everything visible to the model is written to an append-only session log, including system prompts, reasoning, tool calls and results, subagent scheduling, and context injections. A Trajectory view separates records by source. Resume, fork, search, and replay all operate on the same event stream. Agent systems are often difficult to debug because relevant state is scattered across prompts, tool logs, and orchestration code; DeepSeek instead makes the execution history a central data structure.

Four modes expose different operating assumptions. Standard mode provides the full coding-agent toolset, including file editing, shell access, search, skills, planning, goals, subagents, and workflows. Code mode uses model-generated code to coordinate multiple rounds of tool calls. Minimal mode keeps only persistent bash and a str_replace_editor, creating a restrained environment for model benchmarks. Creator mode adds runtime inspection and in-memory Cordis plugin experiments so developers can assemble new presets. A shared kernel can therefore support both product-oriented workflows and narrower evaluation environments.

Composability is the attraction, but it also defines the work left for adopters. Operators need clear rules for plugin permissions, version compatibility, sensitive information in event logs, and the effects of replaying actions against external systems. DeepSeek warns that the project remains a developer preview and that core plugins and APIs will continue to evolve. The source is available under the MIT license, with a web entry point through npx @deepseek-ai/dsh web and a public GitHub repository. The meaningful test is not how many plugins can be mounted; it is whether this structure makes long-running agent behavior easier to understand, reproduce, and safely modify.