Frontend · Backend · GitHub

DeepSeek Harness Turns the Agent Skeleton Into a Fully Pluggable Runtime

By 程序员_小雨 · Aug 14, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Agent frameworks have locked developers into a black-box skeleton where only tools and prompts are user-serviceable. Harness is the first runtime to make the entire agent skeleton pluggable, so teams that need custom execution logic, multi-agent isolation, or deep context engineering can build it without forking and maintaining a private fork of LangChain or the OpenAI SDK.

Summary

DeepSeek Harness treats the entire agent as a plugin assembly. Its Cordis meta-plugin base supports hot-loading, hot-unloading, and automatic state rollback, so the main loop, session storage, task scheduler, and event system are all replaceable without forking source code. A Preset isolation mechanism runs multiple independent agent environments inside a single process, each with its own tools, prompts, permissions, and storage.

An intent-first execution model keeps the LLM focused on outputting core intent while the harness handles tool dispatch, retries, and context assembly, slashing token waste from intermediate chatter. Lifecycle hooks like `agent/pre-step` expose every execution step for interception and context rewriting, with full event tracing for debugging.

It is a v0.1 developer preview with no forward compatibility guarantees, a near-empty plugin ecosystem, and sparse documentation. The architecture targets teams building custom agent platforms, multi-agent systems, or long-lived infrastructure where off-the-shelf frameworks hit a customization wall.

Takeaways

— Every core agent component—main loop, state management, session storage, task scheduling, sandbox, event system, UI—is a swappable plugin with no privileged kernel.

— The Cordis meta-plugin base enables hot-loading, hot-unloading, and automatic state rollback when a plugin is removed, avoiding service crashes.

— Preset isolation runs multiple fully independent agent environments in a single process, each with its own tools, prompts, permissions, and storage.

— An intent-first execution pattern has the model output only core intent; the harness handles execution and returns only key results to the model, cutting token consumption.

— Lifecycle hooks like `agent/pre-step` allow interception and modification of context at every execution step without touching framework source code.

— Full-chain event tracing records every prompt, inference, call, and sub-task transfer for debugging.

— The v0.1 preview is unstable with no forward compatibility; APIs, storage structures, and plugin specs will change, so production use is explicitly discouraged.

— The project is MIT-licensed and targets teams building custom agent platforms, multi-agent systems, or long-lived agent infrastructure.

Conclusions

Harness shifts the agent formula from 'better model = better agent' to 'Agent = LLM + Harness,' arguing that model capabilities have homogenized enough that the runtime base is now the real differentiator.

By making the entire skeleton pluggable, Harness eliminates the fork-and-maintain trap that hits teams the moment their agent logic diverges from what a framework's black box permits.

The Preset isolation model challenges the common workaround of deploying multiple service instances just to run different agent types side by side, which wastes resources and complicates ops.

DeepSeek's strategic move is not another model release but an infrastructure play—if the plugin ecosystem grows, a composable agent base could become a standard layer underneath many AI products.

Concepts & terms

Cordis

The self-developed meta-plugin base underneath DeepSeek Harness that provides spatiotemporal composability: plugins can be hot-loaded and hot-unloaded (time), and they register by key with automatic dependency resolution and loading order (space).

Preset isolation

A Harness mechanism that runs multiple independent agent environments inside a single process, each with its own tools, prompts, permissions, storage, and execution logic, without cross-contamination.

Intent-first execution

An execution pattern where the LLM outputs only the core intent, and the harness handles tool dispatch, retries, and execution, returning only key results to the model to keep context lean and save tokens.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗