I Tried Deepseek's New SuperAgent Harness... Is IT ANY GOOD?

Join this channel to get access to perks: https://www.youtube.com/channel/@aicodeking/join DeepSeek Harness article (Exploring Super-agent Design): https://x.com/DeepSeekHarness/status/2116857257245987236 dsh-superagent prototype (research code): https://github.com/deepseek-ai/dsh-superagent-prototype PandaCode: https://github.com/aicodeking/pandacode Cline Desktop: https://docs.cline.bot/usage/cline-desktop ClinePass: https://docs.cline.bot/getting-started/clinepass In this video, I'll go through DeepSeek Harness's article "Exploring Super-agent Design", which starts from a simple problem: agents can now work on their own, but you can only follow about 3 to 5 sessions at once. I'll explain the design: the four things a Super-agent needs, the Router and Workers split, Triggers with two filters, Chain of Recall, the Watchdog, and what their experiment on synthetic workloads actually shows. Then I'll show how I would build a practical version today with PandaCode bots, using a cheap DeepSeek V4.1 Flash Router, ClinePass models for the Workers, and Cline Desktop for the recurring jobs. Finally, I'll be fair about what this setup doesn't do and how to try the real dsh-superagent prototype safely. -- Timestamps: 00:00 - The 3 to 5 session limit and the Super-agent idea 01:26 - The four things a Super-agent needs 01:56 - Router and Workers 02:58 - Triggers, filters and Chain of Recall 04:25 - Keeping it alive: background commands and Watchdog 05:00 - The experiment (synthetic workloads) 07:02 - Building it with PandaCode bots 09:02 - Models with ClinePass and setup 09:57 - How it looks in Telegram 11:39 - Where Cline Desktop fits 12:17 - Being fair: what this setup doesn't do 13:21 - Verdict