DeepSeek open-sources Harness v0.1, a plugin-built agent framework

DeepSeek has released DeepSeek Harness v0.1 in developer preview, an open-source agent runtime built around a single architectural idea: every part of an agent, from the model to the sandbox to the UI, is a swappable plugin. The release ships alongside DeepSeek-V4-Pro's move out of preview and gives developers a self-hosted alternative to integrated coding-agent products.

What's new

DeepSeek's own product page states plainly: "DeepSeek Harness is now in developer preview for agent harness developers worldwide — source code included." The framework is released under the MIT license and is built on what DeepSeek calls the Cordis kernel, described on the page as a system where "every capability is a plugin that can be swapped or recomposed: models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the UI."

The pitch draws a clear line between model and harness: "The model is the soul of an agent. A harness lets an agent understand its environment, use tools, and keep working in real-world settings." DeepSeek also highlights session-level observability, noting that "everything the model sees is recorded in an append-only session log: system prompts, reasoning, tool calls and results, subagent scheduling, and every context injection" — a design aimed at making agent runs auditable rather than opaque.

Context

Harness arrives the same week DeepSeek shipped DeepSeek-V4-Pro's general-availability release and set a new peak/off-peak API pricing schedule, forming a coordinated push into the agentic-coding market where Anthropic's Claude Code and similar integrated tools currently set the pace. Where those products bundle model, tool orchestration, and interface into one closed system, DeepSeek Harness exposes each layer as a plugin developers can replace independently — swap the model, swap the sandbox, swap the scheduler — while keeping DeepSeek's own models as the default backend.

Why it matters

Open-sourcing the harness layer, not just the model weights, is a different competitive move than DeepSeek's earlier V3 and R1 releases: it invites the developer ecosystem to build on DeepSeek's agent architecture directly rather than just call its API. If Harness gains traction the way DeepSeek's open-weight models did, it could pull developers who might otherwise standardize on a closed coding-agent product into DeepSeek's orbit for the surrounding tooling as well as the model — extending the same low-cost, open-source playbook that has already reshaped API pricing to the agent-framework layer of the stack.