DeepSeek Harness · 2026-08-13

Inside
DeepSeek Harness

A look inside the harness the Chinese company just published

The rules that turn a model into an agent able to work with tools, code, permissions and state.

Start here 5 min Explore the map
  • 49 packages in the repo
  • 88 pages of paper
  • 23 sections
  • MIT license

Overview

DeepSeek has open sourced the layer of software that wraps a model and turns it into an agent that works.

On 13 August 2026 DeepSeek announced two things. One was V4-Pro, its updated flagship model. The other was DeepSeek Harness, and in the medium term it may matter more.

DeepSeek Harness is not a model. It is everything around one: the tools, the files, the terminal, the permissions, the memory of what has happened and the interface. Everything a model needs to stop answering questions and start doing work.

What stands out is how it is built. Every piece is interchangeable, including the one that decides what the agent does next. There is no protected core: if you want to change something, you add a piece alongside the others.

Underneath that idea sits a framework called Cordis, and underneath Cordis sits an 88 page academic paper with formal proofs, signed by researchers from Peking University and DeepSeek. A project publishing the theory that holds it up is not the usual thing.

This site explains the three layers. How it is used, how it is built and what the paper says. You can stay on the first one and walk away knowing how to run it, or go all the way down and understand why someone bothered to formally prove that a program can rewrite itself without breaking.

One warning. DeepSeek Harness is a developer preview, and its authors warn in capital letters that there will be compatibility breaking changes. Everything here is extracted from a specific commit, whose reference and date sit in the footer of every page.

You do not need to read all of it. Choose how deep you want to go.

Reading routes

The map

Twenty-three sections across three depths

Usage 4
02 4 entry points Up and running in 60 seconds npx, source build, headless CLI and Python SDK Read → 03 4 modes The four modes Standard, Code, Minimal and Creator, and what each is actually for Read → 04 3 provider routes Models and providers DeepSeek, catalog providers and any OpenAI-compatible endpoint Read → 05 fail-closed Permissions and sandbox workspace-write, the ask policy and the fail-closed principle Read →
Architecture 8
06 0 privileged core Cordis, the microkernel Services, typed events and revertible effects on a shared context Read → 07 4 layers Profiles, bundles and layers How the tree your machine boots is composed, layer by layer Read → 08 7 core services The plugin tree The core packages and the context key each one registers Read → 09 6 waterfalls Anatomy of a turn From turn/start to turn/end, event by event Read → 10 append-only The session log Model-visible means logged, and what that buys you Read → 11 3 roles Capability seams Interface, provider and consumer, and why one swap moves a whole world Read → 12 18 extension points Where new behavior goes Goal to mechanism lookup, straight from the repo Read → 13 1 patch Write your first plugin A minimal tool, registered, mounted by patch and published Read →
Theory 6
14 87 out of 100 The problem Dynamic composition had no formal foundation, and VS Code proves it Read → 15 every effect, its inverse Revertible effects The temporal dimension: fully undoing what a component did Read → 16 activate / deactivate Reactive coeffects The spatial dimension: declaring dependencies and reacting to their changes Read → 17 5 theorems The calculus and its five theorems Preservation, temporal, spatial, progress and confluence, in plain words Read → 18 4,000 plugins From theorem to code Every symbol mapped to a real API, plus the Koishi case study Read → 19 what exists and what does not Can a harness evolve on the fly? What already works, what the architecture proposes and what it could merely allow Read →
Context 5
20 the central idea Same model, different agent Why two products running the same model behave so differently Read → 21 6 harnesses DeepSeek Harness against the others Against Claude Code, Codex, Pi, OpenHands and opencode Read → 22 4 states What is real Available, evolving, architecture and future, claim by claim Read → 23 TL;DR in 5 points Why it matters What it means for developers, open models, infrastructure and the industry Read → 24 your own endpoint Running it on Helmcode A practical note from the authors of this research, at the end and set apart Read →

What was analysed

Commit analysed
47f9438
Version
0.1.0-rc.5
License
MIT
Extraction date
2026-08-13
Links RepositorioPaperCordisDocumentationPlugins

A Programming Paradigm for Spatiotemporal Composability. Yifan Shi, Wei Zhang, Tianyi Cui. 88 pages, 2026-08-13. preprint en revisión activa.

Extracted from commit 47f9438 on 2026-08-13. DeepSeek Harness is a developer preview and its authors warn of compatibility-breaking changes. The paper is a preprint under active revision.

GitHub figures on 2026-08-14: 78,542 stars, 6,834 forks. They expire.

Independent research by Helmcode. Not affiliated with DeepSeek. Logos and trademarks cited belong to their owners.