DeepSeek open-sources Harness, a plugin-based agent runtime with replayable session logs
- DeepSeek Harness entered developer preview with its source code included, positioning itself as a configurable runtime for developers building AI agent harnesses.
- The runtime uses the Cordis plugin system, where models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the UI are separate plugins that developers can swap or extend through configuration.
- An append-only session log records system prompts, model reasoning, tool calls and results, subagent scheduling, and context injections; its Trajectory view can inspect, resume, fork, search, and replay the same event stream.
- Standard mode includes file editing, shell access, file and web search, planning, goals, subagents, and workflows, while Code mode exposes tools to model-generated TypeScript for multi-step operations.
- Minimal mode limits the agent to persistent bash and a str_replace_editor for benchmarking, while Creator mode adds runtime inspection, in-memory Cordis plugin experiments, and preset-authoring guidance.
Hacker News opinions
I would need a reason to try it over omp. The announcement does not make the comparison clear.
I think the append-only Trajectory view is the most interesting part. Being able to inspect prompts, tool results, scheduling, and context injections, then replay or fork the same event stream, is useful.
I do not see a technical requirement for Node.js here. An agent mostly needs HTTP, tool calls, shell access, and a UI, and there are implementations in Rust, Go, Python, Clojure, and PHP.
Node and TypeScript still reduce extension friction. npm distribution, cross-platform runtimes, async support, types, and the React ecosystem make plugin-heavy tools quick to iterate on.
I use a third-party harness because model choice and visible token spend matter more to me than a vendor pairing. With OpenCode, Luna, and fresh context, I spend about $1-$2 on days I develop instead of subscribing to a $200/month plan.
I found the GitHub README too bare to explain what this actually is. The landing page and generated docs give much better context, while Cordis describes itself as under active development with an unstable API.
I read Harness as the layer behind Claude Code-like agents, not merely a TUI. Cordis appears to support plugins that can be mounted or unmounted at runtime, though tracking inverses in memory may not scale.
I have used my own framework around pi.dev with several models, and vendor harnesses have not felt better adapted. I would rather keep the ability to switch models than rely on a first-party harness.