DeepSeek Harness crosses 100K GitHub stars in 45 hours, "everything is a plugin" architecture sparks Agent framework debate
DeepSeek open-sourced Harness v0.1 developer preview on August 13 (MIT license). It hit 105K GitHub stars and 10K forks in 45 hours, reaching 116K stars by August 16 with over 4,300 third-party plugins within three days.
DeepSeek open-sourced the developer preview of its first Agent product, Harness v0.1, on the evening of August 13 under the MIT license. "Harness" refers to the agent-control software that wraps large models, enabling agents to autonomously break down tasks, call tools, and self-correct. DeepSeek Harness follows the "everything is a plugin" philosophy, making the model, tools, skills, sessions, sandboxes, storage, loops, scheduling, and UI all pluggable and swappable.
Built on the Cordis plugin system, Harness emphasizes traceability: everything the model sees is recorded in an append-only session log, including system prompts, reasoning, tool calls, and sub-agent scheduling, inspectable via the Trajectory view. Harness ships with Standard, Code, Minimal, and Creator runtime modes. Developers can get started with `npx @deepseek-ai/dsh web`.
Harness drew extensive developer testing after release: by August 15, GitHub stars hit 105K with 10K+ forks; by August 16, stars reached 116K, breaking GitHub repo growth records and outpacing R1 and Grok-1 at the same point. Third-party open-source plugins built around Harness exceeded 4,300 within three days.
A DeepSeek-Peking University paper also surfaced, revealing the self-evolution blueprint behind Harness and the Cordis plugin paradigm, validated over four years via the Koishi project. DeepSeek's competition with Anthropic Claude Code and OpenAI Codex has now formally extended to the Agent infrastructure layer.