DeepSeek officially released its first agent product, DeepSeek Harness. Earlier, the V4 Flash release notes mentioned using Harness's Minimal mode as a test framework, hinting at Harness's existence. APPSO obtained beta access before the launch and noted that Harness includes common features of a local agent workbench, but it's not simply a 'DeepSeek version of Codex.'
Harness's uniqueness lies in its 'agent preset' system, which differs from Codex's scenario-based approach by combining tools, prompts, and capabilities through presets. It includes four built-in presets: Minimal mode for simple edits without compression, search, or Skills; Standard mode for everyday coding and project analysis; Code mode for large-scale searches and parallel reads; and Creative mode (cordis) for developing new presets or plugins.
A preset is essentially a configuration document containing system prompts, tools, context compression, and multi-agent capabilities. Users can copy presets or use Creative mode to have the agent create new ones. Harness also allows integration with third-party models, with default providers including Kimi, OpenAI, Anthropic, Google, and nearly 40 others.