AIBRIEF
A rumored 17 project list for the Chinese AI lab's agent platform reportedly skips high star "vase" (showcase/decorative) repos and favors small, mostly unknown infrastructure tools. Vertical applications (industry specific, e.g.
A leaked shortlist of 17 open-source projects tied to DeepSeek's upcoming "Harness" agent platform skews heavily toward small, mostly-unknown infrastructure tools, according to a Leiphone analysis by reporter 高允毅. The Chinese AI lab, the report says, left out the high-star "vase" repos, the flashy showcase projects with thousands of GitHub stars, and picked lean tools built for the work that makes agents actually run.
The pattern: roughly 70% of the selected projects reportedly come from individuals or small teams with near-zero attention, and vertical applications (medical, legal, financial) account for under 4% of the partial list. The biggest shares go to MCP plugins and coding-agent tooling, with secondary emphasis on agent runtime, orchestration frameworks, visual UI, and multi-agent scheduling.
Four projects are named in the report. openma-ai/open-managed-agents is an execution-layer sandbox positioned as a Claude Managed Agents-compatible harness. try-works/role-model is a capability-aware routing protocol with multi-vendor failover. aresbit/MateBot is a mobile remote-control interface for local agents with persistent memory. A fourth project, cut off in the source text, is unrecovered.
The list is unconfirmed, attributed to Leiphone rather than to a DeepSeek announcement. As model-quality gaps flatten, the next moat is the infrastructure that lets agents run safely, route cheaply, and survive long tasks, and star count is no longer the signal to watch.