DeepSeek just dropped something more interesting than another model release. DeepSeek Harness v0.1 is now out in developer preview, fully open-sourced under MIT, and it takes a fundamentally different position on how AI agents should be built. The project, shipped as dsh, frames the whole problem as Agent = Model + Harness — but unlike most frameworks that hard-code their agent loop and tool registry, Harness makes every single component a swappable plugin.
That’s not marketing spin. The README’s opening line is blunt: everything is a plugin. Models, tools, skills, sessions, sandboxes, storage, loops, scheduling, even the UI — all sit behind Cordis plugin boundaries. Developers can swap any of them through configuration without touching Harness source code. The foundation is Cordis, a meta-framework described in a paper called ‘A Programming Paradigm for Spatiotemporal Composability,’ which handles plugin mounting, unmounting, and dependency resolution while keeping capabilities out of any privileged core.
The release ships with four runtime modes. Standard gives you the full coding agent: file editing, shell access, web search, skills, planning, goals, subagents, and workflows. Code mode exposes those tools through a TypeScript SDK so models can combine multi-step operations in a single program. Minimal strips everything down to just bash and str_replace_editor — deliberately bare for benchmarking models in controlled environments. Creator mode adds runtime inspection and in-memory plugin experiments for people building presets.
One detail that stands out: every run is traceable through an append-only session log that records not just tool calls but every context injection — system prompts, reasoning, subagent scheduling, everything the model sees. Resume, fork, search, and replay all operate on that same event stream. Model routing is plugin-based too, with support for DeepSeek, Anthropic, OpenAI, Bedrock, Vertex, Azure, Codex, and any OpenAI-compatible endpoint. Keys are stored write-only in $DSH_HOME/.credentials.yaml. This isn’t a production product yet — the team is explicit about that — but for teams building internal coding agents or benchmarking models in controlled environments, it’s a genuinely different starting point than the fixed-architecture frameworks most people are stuck with.
