DeepSeek Harness: Developer Preview

The DeepSeek Harness (DSH) has officially entered its developer preview phase, offering agent harness creators across the globe a fully composable framework—complete with open source code.

The Core Philosophy: Modular Architecture

At its heart, DeepSeek Harness operates on a simple premise: ~~everything is hardcoded~~ everything is a plugin. This means every single capability can be swapped, modified, or recomposed.

The Agent Equation

To understand the relationship between the intelligence and the environment, we can express it as: $\text{Agent} = \text{Model} + \text{Harness}$

While the model serves as the "soul" or the cognitive engine, the harness provides the necessary sensory and motor functions, allowing the agent to perceive its environment and execute actions in real-world scenarios.

Technical Architecture: The Cordis Kernel

The system is powered by the Cordis kernel, which acts as the central orchestrator.

Key Architectural Features:

Plugin-Based Capabilities: Every function is a plugin. This includes:
- Models & Tools
- Skills & Sessions
- Sandboxes & Storage
- Loops & Scheduling
- User Interface (UI)
Configuration-Driven Composition: Developers can extend or replace any capability via configuration files. This allows for deep customization without needing to modify the underlying DeepSeek Harness source code.

Total Traceability & Observability

DeepSeek Harness ensures that no action is a "black box."

Every interaction the model encounters is captured in an append-only session log. This includes system prompts, internal reasoning, tool invocations, results, the scheduling of subagents, and all context injections.

Users can utilize the Trajectory view to filter and inspect these records by their source. Because everything exists on a single event stream, the following operations are seamless:

Resume a session.
Fork a trajectory.
Search through history.
Replay specific sequences.

Runtime Mode Comparison

Depending on the goal—whether it's benchmarking or full-scale production—DSH offers four distinct runtime modes:

Mode	Primary Purpose	Key Capabilities
Standard	General Purpose	Full toolset: file editing, shell, web/file search, planning, and subagent workflows.
Code	Programmatic Orchestration	All Standard features + Code Mode SDK, allowing the model to write `TypeScript` programs to chain tool calls.
Minimal	Benchmarking	A stripped-down environment containing only a `bash` shell and a `str_replace_editor`.
Creator	Development & Prototyping	Standard features + runtime inspection, in-memory plugin testing, and preset-authoring tools.

Getting Started

You can begin customizing your harness immediately through two primary methods:

1. Quick Start (Web UI)

If you have Node.js installed, you can launch the interface instantly using npx:

$ npx @deepseek-ai/dsh web

2. Installation from Source

For those who wish to dive into the code or contribute to the ecosystem:

$ git clone https://github.com/deepseek-ai/deepseek-harness

Joining the Ecosystem

The DeepSeek Harness is currently in developer preview and is undergoing rigorous testing by the community. The goal is to push the boundaries of artificial intelligence using an open-source infrastructure that is both reusable and composable.