跪拜 Guibai
← All articles
Artificial Intelligence

DeepSeek Harness Turns Every Agent Component into a Swappable Plugin

By 宅小年 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Most coding agents ship as opinionated, monolithic stacks. DeepSeek Harness instead gives developers a plugin architecture where every piece—model, tools, loop logic—can be independently swapped or extended, making it a practical base for custom agent workflows rather than another take-it-or-leave-it product.

Summary

DeepSeek Harness is a newly open-sourced Agent runtime framework built on the Cordis meta-framework. Its core design treats every component—models, tools, skills, sandboxes, file systems, agent loops, and the Web UI—as a plugin. Nothing is hardwired; developers can swap models, add or remove tools, and reconfigure execution loops without rewriting the framework. Plugins collaborate through services and events, and uninstalling a plugin also cleans up its side effects.

The framework ships with four preset modes that demonstrate this composability. Standard mode bundles a full set of capabilities. PTC mode lets the model generate TypeScript to batch multiple tool calls into a single `run_code` invocation, cutting round-trips on long chains. Minimal mode strips down to just a persistent Bash shell and a string-replace editor for straightforward code-editing tasks. Creative mode adds runtime inspection and the ability to create new presets, though it can execute model-generated plugin code and demands a sandboxed environment.

DeepSeek Harness is still in Developer Preview with no compatibility guarantees. It is not yet a drop-in replacement for Claude Code or Codex, but it offers a foundation for developers building custom AI workflows who want control over how their agents operate rather than accepting a fixed agent shape.

Takeaways
DeepSeek Harness is an MIT-licensed Agent runtime framework where models, tools, skills, sandboxes, file systems, agent loops, and the Web UI are all plugins.
It runs on the Cordis meta-framework; plugins communicate via services and events, and uninstalling a plugin removes its capabilities and side effects.
Four preset modes ship out of the box: Standard (full-featured), PTC (batches tool calls via generated TypeScript), Minimal (persistent Bash plus a string-replace editor), and Creative (runtime inspection and preset creation).
PTC mode reduces model-tool round-trips on long call chains by generating a TypeScript program that combines multiple tool calls into one `run_code` execution.
Creative mode can run model-generated plugin code, so it requires an isolated project and a controlled environment.
The project is in Developer Preview; the team warns of potential breaking changes and does not yet recommend it as a replacement for Claude Code or Codex.
Installation is a single npx command (`npx @deepseek-ai/dsh web`) or a standard pnpm build from source, with a Web UI at localhost:3080.
Conclusions

Framing the agent runtime itself as a plugin composition problem, rather than a monolithic application, shifts the competitive ground from "whose agent is smarter" to "whose agent is more malleable."

The four preset modes are not separate agents; they are the same Harness with different plugin assemblies, which demonstrates the architecture's flexibility more clearly than documentation could.

Creative mode's ability to let an agent inspect and modify its own plugin set edges toward self-reconfiguring agents, but the safety warning about running model-generated plugin code signals that the guardrails are still manual.

Concepts & terms
Harness
In the context of AI agents, a harness is the runtime framework that connects a large language model to tools, file systems, sandboxes, and execution loops. The model decides what to do; the harness governs how it does it.
Cordis
A TypeScript meta-framework for building plugin-based applications. Plugins declare dependencies, communicate through services and events, and can be cleanly uninstalled with their side effects reverted.
PTC Mode
A preset in DeepSeek Harness where the model generates a TypeScript program to batch multiple tool calls into a single `run_code` invocation, reducing back-and-forth latency on multi-step tasks.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗