Anthropic details multiagent experiments showing Claude agents can wage a "turf war" over incompatible goals, fail to coordinate, collude on prices, and more (Rebecca Bellan/TechCrunch)

DeepSeek has open sourced the DeepSeek Harness, a new agent runtime for developers. The Node.js-based harness is available on GitHub as a developer preview under an MIT license. Within just a few hours, the repository gained over 33,000 GitHub stars, indicating significant interest. The DeepSeek Harness stands out because "everything is a plugin," including the model adapter, tool registry, session log, and agent loop.

Each plugin is replaceable, and there is no privileged core to patch, allowing developers to extend the harness by mounting a plugin alongside the others. The architecture is based on Cordis, a meta-framework for spatiotemporal composability. The harness ships with four presets: Standard, Minimal, Code, and Creator. Standard mode provides a full coding agent with various capabilities, while Minimal offers only two basic tools.

Code mode generates a TypeScript SDK for model interaction, and Creator mode enables developers to create custom agent presets. The harness also logs every model request in an append-only session log, ensuring transparency and enabling features like resume, fork, replay, transcripts, and telemetry. The local backend wraps subprocesses in Linux Landlock, macOS Seatbelt, or a Windows ACL restricted-token runner to provide sandboxing.

The harness supports multiple inference providers, including DeepSeek, Anthropic's Claude Code, and OpenAI's Codex. However, the project currently does not accept external pull requests, instead inviting contributions through GitHub Discussions and plugin development. DeepSeek sees its plugin architecture as a major differentiator, despite the presence of other open-source harnesses in the market.