DeepSeek unveils AI agent ‘Loop’ design

A developer gives an AI an instruction. The AI edits files, executes computer commands and searches the web before reviewing the results again. DeepSeek has released a program that bundles this process together, the ‘DeepSeek Harness,’ alongside its latest V4-Pro model. It is open source, so anyone can download and modify it.

The tools connected by the harness include file editing, computer command execution, web search and calls to auxiliary AI systems. The program organizes the conversation and manages the work in a sandbox, a secure execution environment. It is less a system that simply produces answers than a coordinating layer that enables the model to continue carrying out real tasks.

DeepSeek calls its approach ‘loop engineering.’ Whereas prompt engineering involves refining a single question or instruction, loop engineering determines when to perform the next task, when to verify the result and when to stop. dsh separates this loop from the model, conversation session and execution environment, turning each into an independent component that developers can replace wholesale and test.

What, specifically, changes as a result? Without changing the model, developers can modify the task sequence and verification method to experiment with reducing problems such as agents repeating the same task or continuing to push through an incorrect result. DeepSeek has applied the MIT license, opening the project to broad use, including commercial applications.

The release is also an attempt to establish a standard for the ecosystem. If the loop design becomes a widely used framework, subsequent tools and benchmarks may use its structure as a reference. However, major agent tools such as Anthropic’s Claude Code and OpenAI’s Codex keep their loop designs private, so it will only become possible to assess where control lies after developers have accumulated experience using and improving them.