AI Pulse · Under the Radar

DeepSeek Harness has become the fastest-growing repository on GitHub today, reaching 206,307 stars in 18 days. That is roughly 11,274 new stars a day. It is a plugin-based, model-agnostic runtime for running AI agents, meaning it is designed to let developers connect different models, tools and integrations rather than being tied to one provider.

But the conversation around it is changing. Early attention focused on the speed of its adoption. Now developers are testing it in real workflows and raising practical concerns. Reports point to token use around 10 times higher than comparable agent harnesses, a bug that duplicates context, and a much lower working-integration rate than the headline number suggests. Of 260 integrations flagged, only 41 were reportedly working. One developer described it bluntly: “this isn’t another Codex, it’s a breadboard.”

That distinction matters if you run a company. A fast-moving open-source project can be useful for experimentation without being ready for a business-critical process. High token usage becomes a real operating cost. Broken integrations create manual work. Context problems can also lead to inconsistent outputs and harder-to-trace failures.

The sensible move is to separate interest from readiness. Watch projects like this, test them in contained use cases, and measure cost, reliability and human intervention before putting them near core operations. This is the kind of thing we build into an AI command centre, where tools are assessed against the work they actually need to support. The project’s current activity and discussions are visible in the DeepSeek Harness repository.