DeepSeek Made the Agent Loop Itself a Plugin — 196k Stars in 13 Days
DeepSeek's agent harness passed 195.9k stars in under two weeks on one idea: every part is swappable, including the loop. The architecture is right. The warning printed next to it is the part to read.
NeuroX AI · August 26, 2026
DeepSeek open-sourced its agent harness on 13 August. Thirteen days later the repo sits at 195.9k stars and 22.2k forks.
The tagline is the architecture: Everything is a Plugin. The official page lists what's swappable — "models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the UI." Read that list again. loops is in it. The agent loop — the thing that decides when to call a tool, when to stop, when to hand back — is a mountable component, not a core you patch around.
That's the right instinct, and we have numbers for why. Same model, different harness, three points of swing on the same benchmark. If the harness is the variable that moves your results, it should be the part you can change.
But swappable is not stable. DeepSeek says so plainly: the project "remains in developer preview" and "its core plugins and APIs will continue to evolve." There are already 13,147 commits on master. With no privileged core to anchor against, every interface you build a plugin against is a moving target — and the failure won't look like a crash, it'll look like a loop that quietly stops calling your tool.
Fine trade for exploration. Bad trade for a system on call. Pin a commit, wrap the one plugin boundary you actually depend on, and treat the rest as someone else's roadmap.