deepseek harness

DeepSeek Harness (often shortened to dsh) is an open-source, MIT-licensed agent runtime framework developed by DeepSeek AI. It is designed to bridge the gap between a language model and the real-world environment, enabling the model to function as an autonomous agent capable of executing multi-step tasks.

Core Concept: "Agent = Model + Harness"

DeepSeek frames the architecture of an agent as a combination of two distinct parts:

• The Model: Handles reasoning, logic, and content generation.
• The Harness: The runtime layer that provides the "body" for the model. It supplies the tools, session management, sandboxing, and execution loops required to interact with a file system, terminals, and external APIs.

Key Features

• Plugin Architecture: Built on the Cordis framework, DeepSeek Harness utilizes an "everything-is-a-plugin" design. Components such as the model adapter, tools, session store, sandbox, and the agent loop itself are modular. This allows developers to swap, extend, or recompose capabilities through configuration without modifying the core source code.
• Agent Runtime: It manages the state and persistence of tasks, supporting features like resume, fork, search, and replay via an append-only event log.
• Modes of Operation: It supports various runtime modes to suit different needs, such as:

* Standard: Full toolset for general agentic tasks.

* Minimal: A stripped-down mode typically used for benchmarking.

* PTC/Creator: Specialized modes for orchestration code generation or experimenting with new plugins.

Important Distinction

It is important not to confuse DeepSeek Harness with the Language Model Evaluation Harness (often associated with EleutherAI), which is a common tool used by researchers to benchmark the performance of language models (like DeepSeek-V3) on standardized tests like MMLU or GSM8k. While DeepSeek models are frequently evaluated using the latter, "DeepSeek Harness" refers specifically to the agent runtime framework.