DeepSeek Harness modes: Standard, Code, Minimal, Creator
Standard, Code, Minimal, and Creator — what each official DeepSeek Harness runtime mode is for.
Last updated: 2026-08-14
From the official landing page at deepseek.com/harness. Unofficial restatement.
The formula they published: Agent = Model + Harness. The harness is a Cordis kernel plus a pile of plugins. A "mode" is just a preset of those plugins.
Standard
Full coding agent. File editing, shell, file and web search, skills, planning, goals, subagents, and workflows.
Use this if you want the default product.
Code
Everything in Standard, plus the Code Mode SDK. The model writes TypeScript to combine several tool calls in one program instead of chatting through every step.
Use this for long, mechanical multi-step work.
Minimal
Two tools: persistent bash and str_replace_editor.
This is the thin environment DeepSeek used for public Code Agent benchmarks (the "极简模式" mentioned in the V4-Flash API notes).
Use this to bench the model, not to daily-drive.
Creator
Standard plus runtime inspection, in-memory Cordis plugin experiments, and preset authoring.
Official messaging: you can inspect the current runtime and compose new modes. Community reports say dynamically generated plugins currently live in memory and vanish on restart.
Use this if you are building harnesses, not just using one.