Summary
DeepSeek has released Super Agent, a research prototype that some are calling its answer to Claude Code. Instead of waiting for a developer to open a session, it watches a team’s GitHub repo and chat tool, picks up issues and pull requests when authorized people ask, works in its own session and directory, and replies with a finished PR. This video from Sam Witteveen looks at what it does and, more importantly, how it was built.
The key is the DeepSeek harness, released in August and now one of the most-starred GitHub projects this year. Super Agent is essentially a set of plugins, configured through a layered patch system with bundles, profiles, and command-line overrides. The video explains the harness’s reversible, composable plugin design, its four UI modes, and a creator mode where the agent writes and installs its own plugins.
The presenter also asks whether the model or the harness matters more, citing guests from Instacart and Box who favored the harness. He points to the Frontier Harness Eval, which ran one model, Kimi K3, through nine coding-agent harnesses on the same 30 tasks. Differences in results show how much the surrounding harness shapes outcomes.
📺 Source: Sam Witteveen · Published November 03, 2026
🏷️ Format: Deep Dive
