AI Briefing
KO
Pick

DeepSeek: Plugin-based Agent

·2026.08.16 15:30

Key point

DeepSeek Harness, released by DeepSeek AI, is a modular agent framework where even the agent loop can be replaced via plugins.

Details

DeepSeek Harness (dsh), released by DeepSeek AI, is an agent framework designed around the philosophy that "everything is a plugin." While existing coding agents provided APIs to a fixed core, dsh adopts a plugin tree structure assembled at boot time, allowing users to replace or patch any component, including the agent loop itself.

This system is based on the Cordis framework and mounts model adapters, tool registries, session logs, and more as plugins without modifying the core. It operates based on effects registered by plugins, featuring a rollback structure where related registrations are reverted when a plugin is removed.

Key features and components include:

  • Profile and Bundle: Configuration units that hierarchically stack settings and code.
  • Capability Seam: Separates service definitions, providers, and consumers to modularly extend features such as filesystem access or process isolation.
  • Session Log: Records all model context to ensure runtime immutability and allow reconstruction of all requests.

Currently in Developer Preview stage, frequent breaking changes may occur, so it is recommended to use it in isolated environments such as containers for security.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.