Ask a machine how it feels, and watch it answer
Charts in chat, ten-second live sampling with per-process attribution, and a real bandwidth test. Every one of them reads. None of them write.
Until this week, asking Wegweiser about a machine's performance got you an essay. Accurate, but an essay. You would ask how the CPU had been behaving and receive a paragraph of prose, or worse, a table of eight hundred data points that nobody was ever going to read.
Three things shipped that change the shape of the answer.
Charts, drawn from the data rather than described
Wegweiser's assistant now renders charts directly in the conversation. Ask for a health trend, a day of CPU history, or a live sample, and the numbers arrive as a chart in the reply.
The mechanism matters more than the pixels. The AI does not draw the chart. It receives the summary statistics, the trend, the extremes, the anomalies; the chart itself is computed from the underlying series by the platform and rendered deterministically. The model narrates what the chart shows, and the two cannot disagree, because the model never had the opportunity to redraw the data. If it says "a spike to seventy-five percent", the spike is in the picture, at seventy-five percent, because both came from the same measurement.
Charts persist with the conversation, so a thread you reopen next week still shows what you were looking at.
Ten seconds of live truth
The new live sampling verb watches an endpoint in real time for up to fifteen seconds and reports what actually happened during the window:
- CPU and memory, sampled several times a second, with the processes that consumed the CPU named alongside and their share of total capacity.
- Network throughput, down and up, on a single chart.
- Disk throughput, IOPS and queue depth, plus which processes did the reading and writing.
Queue depth is the one to watch. A machine can look calm on a CPU graph while its disk queue is backing up, and that is the pattern behind most "everything is slow but nothing is wrong" tickets. Now you can ask, and see.
In a live test on an Ubuntu workstation this week, the disk sample showed a brief read spike and the assistant attributed it, by name, to a security audit that was running at that moment. That is the difference between a monitoring graph and an answer.
Ask it in plain language. "What is eating the CPU right now", "is the disk thrashing", "watch it live for ten seconds" all work.
A real bandwidth test
Wegweiser can now measure the actual throughput of an endpoint's connection: download, upload, and latency, with the ramp charted so you can see the connection reach its ceiling rather than trusting a single number.
Two targets are available. The default measures the internet connection from that endpoint. The second measures the path to the Wegweiser management server, which is the right test when the question is "why is this agent slow to report" rather than "what is this office's bandwidth".
Because a speed test moves real data across a client's link, and client links are sometimes metered, the assistant runs it only when you explicitly ask for one, and tells you how much data it will move before it starts.
What none of this does
Every feature above reads. None of them write.
Wegweiser's assistant cannot install, restart, delete, configure or execute anything on an endpoint. It has no verb that takes a shell command, and no verb that accepts a URL. The bandwidth test is a good example of the discipline: its measurement endpoints are compiled into the agent itself, and the only parameters the platform can send are sizes. A speed test that accepted arbitrary destinations would be a data exfiltration tool with a friendly name, so ours does not accept them.
The limits are enforced twice, once by the platform and again by the agent on the endpoint, which holds its own copy of the rules and refuses anything outside them. A compromise of the web application still cannot make your fleet do something the agent does not already permit.
When something genuinely needs changing, you get a terminal and a human presses enter.
Availability
The features are live for all tenants. The endpoint capabilities arrived with agent versions 0.3.84 through 0.3.86, which roll out automatically through the update channel; devices update themselves within the hour and older agents simply do not offer the newer tools until they do. Nothing to install, no maintenance window.
If you want to try the whole set in one sitting, open any device, ask it what is eating the CPU right now, then ask it to run a speed test. That is the shortest path to seeing what changed.