DeepSeek Harness

"Everything is a plugin." DeepSeek's open-source agent harness (dsh), the execution layer it uses to evaluate and deploy its own models and the scaffold behind the V4.1-Flash agentic results. Built in TypeScript on the Cordis plugin kernel, which only loads, unloads, and resolves dependencies between plugins; every agent capability — model adapters, tools, skills, sessions, sandboxes, storage, the agent loop, scheduling, and the UI — is a plugin that developers swap or extend from configuration without touching the source. Everything the model sees is written to an append-only session log (system prompt, chain of thought, tool calls and results, sub-agent dispatch, every context injection) and surfaced in a Trajectory view; resume, fork, retrieval, and replay share that one event stream.

Ships four presets: Standard (a full coding agent with file editing, shell, file and web search, skills, plans, goals, sub-agents, and workflows), PTC (the same tools exposed through a Code Mode SDK so the model composes multi-step actions in one TypeScript program), Minimal (a two-tool bash plus str_replace_editor agent for benchmarking models in bare environments), and Create (runtime inspection and in-memory plugin experiments for authoring new presets); an Agent Team mode runs a lead agent over sub-agents. Launched as a developer preview in mid-August 2026 with the repository public from August 13, installable as npx @deepseek-ai/dsh web; at filing it had passed 218K GitHub stars with release candidates landing daily. MIT license; DeepSeek warns of compatibility-breaking changes during the preview.