DeepSeek open-sources Harness, a plugin-driven agent runtime, in developer preview

DeepSeek has released a developer preview of DeepSeek Harness, an open-source framework for building agent harnesses — the layer that lets a language model perceive its environment, call tools, and keep working across real-world tasks. The design principle is that every capability is a swappable plugin: models, tools, skills, sessions, sandboxes, storage, control loops, scheduling, and even the UI. It is built on the Cordis plugin system, whose services and events wire the pieces together, so developers can select, replace, or extend any component through configuration without touching the core source. Getting started is a single npx command or a clone of the public GitHub repository.

A central selling point is full traceability. Everything the model sees — system prompts, reasoning traces, tool calls and their results, subagent scheduling, and every context injection — is written to an append-only session log. A Trajectory view lets developers inspect these records by source, and resume, fork, search, and replay all operate over the same event stream, which is a meaningful debugging and reproducibility story for agent development. The harness also ships several runtime modes: a full-toolset Standard mode, a Code mode that has the model generate code to orchestrate multiple rounds of tool calls, a stripped-down Minimal mode (shell plus file editor) aimed at benchmarking raw model capability, and a Creator mode for inspecting the live runtime and composing new modes from in-memory plugins.

The significance is less the individual features than the packaging: a major model lab is shipping its agent infrastructure as reusable, composable open source rather than a closed product. That positions Harness as competition to existing agent frameworks and invites a third-party plugin ecosystem. DeepSeek cautions that this is an evolving preview and that core plugins and APIs will change, so it is aimed at framework builders experimenting on the tooling rather than teams seeking a stable production platform.