跪拜 Guibai
← All articles
DeepSeek · GitHub · Agent

DeepSeek Harness Lands: A Plugin-Native Agent Runtime That Treats Everything as Swappable

By 程序员晓凡 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

DeepSeek Harness is the first major Agent runtime to make the entire stack—including the Agent loop—a hot-swappable plugin, which means teams can fork and customize their coding agents without waiting on a vendor's roadmap. The 115,000-star launch signals that a large developer base is ready to treat Agent infrastructure as composable rather than monolithic, and the trajectory observability gives debugging and compliance a concrete tool that most current coding agents lack.

Summary

DeepSeek Harness is an open-source Agent runtime that completes the company's earlier model releases with a configurable execution environment. Its core architecture rests on the Cordis microkernel, which handles only plugin loading and dependency management, letting every other component—models, tools, skills, sandboxes, UI, and the Agent loop—be swapped at runtime without crashing state. The framework ships with four run modes (Standard, PTC, Minimal, Creative) that differ only in which tool plugins are loaded, and a trajectory system that logs every event the model sees, filterable by source and forkable into new sessions.

Hands-on testing across six cases shows Standard Mode performing at top-tier Agent levels on Windows: repository understanding in 13 seconds, a coding-and-shell closed loop that self-diagnosed a sandbox denial and escalated privileges, and a refactoring task that dispatched parallel sub-agents to produce a 25-test Python package. Minimal Mode fails outright on Windows because its bash tool does not support win32, and the web retrieval case exposed a missing page-fetch tool that caused non-convergent searching. The plugin ecosystem already counts over 700 community plugins, installable by pasting a GitHub URL into the chat, and the Creative Mode lets users describe a plugin in natural language and have the Agent write, review, and publish it without opening an IDE.

Cost for all test tasks totaled under 8 RMB with cache hit rates above 99%. A price increase took effect August 17, raising V4 Pro peak output to 27 RMB per million tokens, but the platform remains roughly one-tenth the cost of GPT-5.6 Sol for equivalent token volumes.

Takeaways
DeepSeek Harness is built on the Cordis microkernel, which only manages plugin loading and dependencies; every other component is a swappable plugin.
The framework defines Agent = Model + Harness, where Harness is the body that connects the model to filesystems, terminals, web, and toolchains.
Four run modes—Standard, PTC, Minimal, Creative—differ only in which tool plugins are loaded for the session.
Standard Mode completed repository analysis in 13 seconds with accurate line numbers, a coding task with self-verification and sandbox escalation, and a refactoring task using parallel sub-agents that produced 25 passing pytest cases.
Minimal Mode is broken on Windows because its bash tool returns 'terminal inspection is unsupported on platform win32.'
Web retrieval found the official pricing page but could not fetch page bodies; the toolset lacks a fetch tool, causing non-convergent searching.
The trajectory tab logs every event by source (SYSTEM, USER, CONTEXT, TOOL, THINK) and supports forking conversations without polluting the original session.
Community plugins exceed 700; installation requires pasting a GitHub URL into the chat, and Creative Mode lets users describe a plugin in natural language for the Agent to build, review, and publish.
All hands-on tasks cost under 8 RMB total with cache hit rates above 99%, but a price increase on August 17 raised V4 Pro peak output to 27 RMB per million tokens.
Tasks can appear stuck while the page timer keeps running; checking the terminal for the dsh web process is the first debugging step.
Conclusions

Making the Agent loop itself a plugin is a genuinely unusual architectural choice that most Agent frameworks avoid because it complicates state management; pulling it off with hot-swapping suggests the Cordis microkernel is doing heavy lifting on isolation.

The sandbox interception-to-escalation chain observed in testing—where the Agent diagnosed a denial, requested permissions, and triggered manual approval—is a workflow that closed-source coding agents rarely expose to users, and it worked end-to-end without human scripting.

Minimal Mode failing on Windows is not a minor bug; it means the 'baseline' mode for benchmarking is Linux-only, which undercuts reproducibility claims for a large portion of the developer base.

The web retrieval failure is a toolchain gap, not a model gap: the Agent located the correct official URL and dispatched sub-agents, but without a fetch primitive it could only triangulate from search snippets, a limitation that will hit any research-heavy workflow.

Publishing a plugin from a natural-language description to a tagged GitHub repo without opening an IDE collapses the tool-building feedback loop to minutes, which could produce a plugin ecosystem that grows faster than VS Code's extension marketplace did in its early days.

The price increase is steep in percentage terms but the absolute cost remains an order of magnitude below GPT-5.6 Sol, which means the economic pressure on Western API providers is not going away.

Concepts & terms
Cordis microkernel
The minimal kernel at the core of DeepSeek Harness that handles only plugin loading, unloading, and dependency management. It enables hot-swapping any component at runtime without crashing the Agent's state.
Agent = Model + Harness
DeepSeek's official framing: the model is the reasoning engine, while the Harness is the body that connects it to filesystems, terminals, web, and toolchains. Tools like Claude Code and Codex are Harnesses under this definition.
Trajectory
An append-only event log in DeepSeek Harness that records everything the model sees, categorized by source (SYSTEM, USER, CONTEXT, TOOL, THINK). Supports filtering by source and forking conversations from any point.
Run modes (Standard / PTC / Minimal / Creative)
Four session configurations that differ only in which tool plugins are loaded. Standard includes file editing, shell, search, skills, sub-agents, and workflows. PTC adds programmatic tool calling. Minimal has only bash and a file editor. Creative adds plugin inspection, testing, and creation tools.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗