DeepSeek Harness With Local Models: Live Test
I test the DeepSeek Harness on local models through Ollama. The Trajectory view shows every agent step. The web search test timed out because no search tool was set up.
The main idea
DeepSeek launched its own agent harness, and it reached about 130,000 GitHub stars in 3 days. In this live stream I tested it against local models instead of their API.
Setup is one command. Ollama launched DeepSeek Harness, and I used Qwen 3.8 27B MLX. The harness is a web UI. Claude Code and Codex started in the terminal, and DeepSeek chose the browser. I like the Trajectory tab, because it shows each step the agent took, including time to first token. The settings have presets for standard, code, minimal and creator modes. Permission defaults to workspace write.
The tests did not go well on my machine. My Minecraft clone task stalled, and OBS slowed my local models. I switched to Gemma 4 E2B, then asked the harness to search the web. The search tool timed out because I had not set up a search API key. Hermes agent can write its own script to search the web when it has no search tool. Part 1 ends before a verdict, so I will keep testing.
Key points
- Ollama launches the harness with one command, and it installs the harness if you do not have it.
- The harness is a web UI with a Trajectory tab that shows every step the agent took and the time to first token.
- Presets are standard, code, minimal and creator mode; permission defaults to workspace write.
- Web search timed out because no search API key was set. Hermes agent can write its own script to search instead.
- Qwen 3.8 27B added health checks without being asked, so be explicit about what you do not want.
Chapters
- 0:45Ollama launch and setup
- 6:50Web UI and Trajectory tab
- 8:46Settings and agent presets
- 25:22Web search test
- 36:01Switching to DeepSeek V4 Flash
Save my free seat →