Alibaba Cloud's community blog carried a post on 19 October from the OpenAnolis community describing SkillHub, a platform for packaging operations knowledge as AI-agent skills aimed at the infrastructure layer. The community says several hundred skills have been submitted and are being published in batches, and the post features one: AnolisOS-SysDiag, written by forward deployed engineer Wang Yangke, which takes a single 'diagnose this machine' instruction and returns a report with a root cause and a suggested fix.
The stated problem is that agents handle application-layer tasks but stall at the operating system, where a human runs top, iostat, vmstat, free and ss in sequence and correlates the outputs from experience. The skill runs the diagnostic commands, joins CPU, I/O and memory readings, and attaches remediation guidance that accounts for Anolis-specific kernel work — Group Identity scheduling, the SLO scheduler and cgroup v2 changes — because the author argues generic Linux tuning advice can be wrong on that kernel. The piece doubles as a submission call: the community and AMD are running a skill-creation camp on operating-system operations.
Our takeThe claim worth testing is not that an agent can run vmstat; it is that a skill becomes a versioned distribution format for procedure — which commands to correlate, which sysctl value fits which kernel build — so the model calls a reviewed artifact instead of improvising. That works where the tool surface is closed and outputs are parseable, exactly the disk-full-versus-iowait case in the post. Whether the platform's several hundred skills extend past such bounded cases is not shown.