Running DeepSeek Harness locally against Qwen3.8-27B

Spent the last couple of evenings running DeepSeek Harness (dsh) locally against Qwen3.8-27B on my own GPU, and it's the most fun I've had learning agent internals in a long time.

Quick context if you missed it: DeepSeek open-sourced dsh on Aug 13 and it became one of the fastest-starred repos GitHub has ever seen, over 20k stars in the first hour. The core idea is that everything is a plugin. The model adapter, the tool registry, the session store, the sandbox, even the agent loop itself. So you can swap out the model layer, point it at any OpenAI-compatible endpoint, and run the whole harness against a local model. Qwen3.8-27B dropped a day later and fits on a 32GB card, so that's what I'm feeding it.

The part teaching me the most is the session log. Every run writes an append-only event stream: system prompts, the model's reasoning, tool calls and results, subagent scheduling, every context injection. The Trajectory view lets you inspect all of it by source, so I can finally watch what a multi-agent loop actually does step by step instead of guessing from the final answer. And since resume, fork and replay all share that same event stream, when a run fails I can replay it and pinpoint exactly which tool call or context injection went wrong, and where in the sequence.

If you're designing agentic workflows and the orchestration still feels like a black box, this is a genuinely good way to crack it open.

It's still a v0.1 developer preview, so expect some rough edges. However it looks impressive, and it should only get better from here.

Can't wait for the weekend to dig deeper.