How DeepSeek Harness Turns Every Agent Component into a Plugin

DeepSeek recently released an early version of Harness, an open‑source agent framework built on the Cordis meta‑framework. The core idea is that every element of an AI agent—model, tool, skill, session, sandbox, file system, loop, orchestration and user interface—is implemented as a plugin that can be mixed, matched, replaced or extended without changing the core code.

https://github.com/deepseek-ai/deepseek-harness

The term “harness” is used this year to describe middleware that manages prompts, context, tool orchestration, agent loops, state, error handling, security and permissions for AI models.

Other model‑specific frameworks include Anthropic’s Claude Code for Claude models, OpenAI’s Codex for GPT models, and community projects such as Aider, Cline, Goose, OpenCode, OpenHands and Pi. The word “harness” can refer narrowly to the agent loop and tools or more broadly to coordination of sandboxes, sub‑agents and other services, as illustrated by Google’s Antigravity Agent Runtime.

AI model frameworks have become a key competitive arena; their user interfaces create inertia that makes switching costly. Design choices also affect model performance and cost. For example, the Pi coding agent uses a system prompt of about 200 tokens, whereas Claude Code requires roughly 10,000 tokens, leading to different results for the same underlying model.

DeepSeek Harness stands out by leveraging Cordis’s plugin system, which allows dynamic addition and removal of components without causing chaos. The framework supports plugin dependencies and defines two forms of composability: time composability (removing a component without side effects) and space composability (managing dependencies between components). These concepts are detailed in a paper by Yifan Shi, Wei Zhang and Tianyi Cui (see the linked PDF).

https://github.com/cordiverse/paper/blob/main/paper.pdf

The authors compare this to Microsoft Visual Studio Code’s extension system, where extensions run in a shared “extension host” process and cannot be removed without restarting the host. Unlike VS Code, DeepSeek Harness supports explicit plugin dependency declarations.

Time and space composability are deemed crucial for systems that can be modified with minimal manual intervention, avoiding forced restarts or crashes when components appear or disappear.

Harness also provides chain‑of‑thought tracing: every event—system prompts, reasoning steps, tool calls, sub‑agent scheduling and context injections—is recorded in an append‑only session log. The trace view lets users filter by source and supports restore, branching, search and replay based on a single event stream.

DeepSeek’s R1 model, released last year, demonstrated chain‑of‑reasoning by breaking prompts into a series of “thoughts” and reflecting on each step before producing a final answer, a capability useful for evaluating reasoning quality and output accuracy.

In the broader ecosystem, Anthropic now deploys a classifier to detect chain‑of‑thought blocks used for training data, while OpenAI hides raw reasoning for its o1 model, citing user experience and competitive considerations. With the recent launch of DeepSeek‑V4‑Pro and V4‑Flash, the API enables mind‑map mode by default, positioning traceability as a differentiating advantage.

Armin Ronacher, co‑founder of the AI company Earendil, praised the open‑source nature of DeepSeek Harness, noting its inspirational impact on their own Pi‑agent development.