DeepSeek's Agent Harness Hit 161k Stars in Six Days. Don't Rebuild Your Stack Around It.

DeepSeek shipped DeepSeek Harness on August 13. Six days later it sits at 161,000 stars and 16,900 forks, one of the steepest adoption curves GitHub has ever recorded for a dev tool.

It earns the attention. It does not earn a migration.

The architecture is the interesting part. dsh runs on Cordis and takes "everything is a plugin" literally. The model adapter is a plugin. So are the tool registry, the session log, the sandbox, and the agent loop itself. There is no privileged core to fork; you extend it by mounting a plugin beside the others, and every registration unwinds when its plugin unloads.

That is cleaner than what most harnesses ship with. It is MIT-licensed, free, and still flagged developer preview with breaking changes promised in the README.

The star count measures enthusiasm, not durability.

1. A fully swappable harness is an argument that harnesses don't matter. DeepSeek made every layer replaceable from config. That is a bet that no single orchestration loop is worth defending, and the market already voted the same way. Teams have been cycling between Claude Code, Goose, Aider, and OpenCode with close to zero switching cost, and Spotify went through three of them.

2. Your real switching cost sits above the harness. Specs, evals, context conventions, tool contracts, review gates. None of that ports as a plugin config, all of it takes months to build, and it survives whichever loop you're running this quarter. When the harness is a weekend swap, the artifacts that outlive the swap are the asset.

3. Preview software is not a platform decision. The repo is past 12,000 commits and iterating hard. Read it, steal from it, run it on a side project. Standardizing your team on it this month means eating breaking changes you were explicitly warned about.

Better orchestration also leaves your actual constraint untouched. Across more than 25 studies and telemetry sets through mid-2026, feature branch throughput is up 59% year over year while median main-branch throughput fell. Agentic PRs wait 5.3x longer for reviewer pickup. Faros measured median review time up 441.5% across 22,000 developers even as task throughput climbed.

Teams are stuck at review, where a human has to decide whether generated code does what anyone asked for. A written spec makes that call fast, because it turns review from "is this good code" into "does this match what we agreed." A plugin tree cannot do that no matter how cleanly it composes.

So clone it. The Cordis composition model is worth an hour and it will make your own tooling better.

Then spend the rest of the week writing down what you want built, because no harness is going to infer that for you.