DeepSeek's agent harness tops 190,000 GitHub stars in 11 days

DeepSeek released DeepSeek-Harness on August 13. The repo passed 190,000 GitHub stars by August 24, up from about 173,700 on August 20, and the growth hasn't leveled off. Most agent frameworks bundle their model routing, sandbox, and scheduler into one fixed stack, so replacing any single piece means migrating the whole thing.

The project's mounting layer, called Cordis, treats that bundle as optional. Models, tools, skills, sessions, sandboxes, storage, the agent loop, scheduling, and UI each mount independently. A team can swap one layer without touching the rest.

That design matters most for teams whose model or sandbox provider keeps changing. A locked-together framework means a full migration each time a vendor gets swapped; a mounted one means changing a single layer. The repo doesn't say who's running Cordis in production, or what happens when two mounted layers disagree on a session format. Star counts measure attention, not adoption, and an 11-day-old project hasn't proven the pluggable design holds under real workloads.