News · Medium impactBack to News

DeepSeek Open-Sources Modular Agent Framework DeepSeek Harness

DeepSeek has open-sourced DeepSeek Harness (dsh), a modular agent framework. Built on the Cordis meta-framework, the system employs a plugin-first architecture to decouple model adapters, execution loops, and tool registries while providing both CLI and local Web interfaces for transparent and extensible agent development.

Core Positioning and Architectural Philosophy

DeepSeek has officially released the open-source agent framework DeepSeek Harness (dsh), positioned as an extensible agent development and execution environment for coding and complex workflows. The system introduces a design paradigm separating model intelligence from peripheral tool harnessing.

The system embodies the everything is a plugin philosophy, abstracting tightly coupled model adapters, execution loops, tool registries, filesystems, and user interfaces into swappable plugins that developers can configure on demand.

Spatiotemporal Composability Powered by Cordis

DeepSeek Harness is fundamentally built on the spatiotemporal composability of the Cordis meta-framework, supporting non-intrusive mounting and hot-swapping in space, while ensuring that registered services and event listeners are fully revoked upon plugin unloading over time to prevent state contamination during long-running sessions.

Plugin Lifecycle and Execution Flow

  1. Load Config and Plugins

    Reads the declarative cordis.yml topology and instantiates model adapters and tools.

  2. Initialize Execution Loop

    Mounts workspace context, initializes append-only logs, and applies permission controls.

  3. Dispatch Actions and Tool Calls

    Drives LLM reasoning, dispatching file edits, shell commands, and subagent tasks.

  4. Unload and Cleanup State

    Unwinds side effects and releases system resources safely upon task completion or plugin swap.

Key Capabilities and Interface Support

Regarding model compatibility, DeepSeek Harness maintains a model-agnostic design, supporting Anthropic, OpenAI, and OpenAI-compatible third-party APIs alongside DeepSeek models to give developers flexible options for diverse task requirements.

The system adopts a local-first architecture where all session logs and workspace data reside locally. Beyond the terminal CLI, developers can also inspect session histories, review reasoning traces, and approve sensitive actions by launching a local Web console with one command.

BASH
# Launch local web interface
npx @deepseek-ai/dsh web

For complex engineering tasks, the system fully supports repository exploration, precise file editing, shell command execution, web search, and subagent delegation, while recording append-only session snapshots that enable forkable replays from historical nodes.

Developer Preview Status and Security Notes

DeepSeek Harness is currently in Developer Preview, with underlying interfaces and configuration schemas evolving rapidly, meaning future versions may introduce breaking changes.

Next step

Keep tracking DeepSeek Harness

Continue along the same topic.

Open entity record