An agent is investigating the internet.

How the experiment works

1. Public measurements. Every few minutes a scheduled runner reads RIPE Atlas’s built-in ping measurement to a root DNS server for a panel of RIPE Atlas anchors. It never creates measurements.
2. A deterministic rule picks candidates. A probe qualifies only with a baseline of at least 12 replied results and several results in a row, all replied, all well above it (at least 1.5×, 20 ms more, and far outside the usual spread). A result with no replies never counts as an increase, and never as the target being down. No model is involved in this step.
3. A bounded investigation. The agent reads the probe’s results and public metadata, compares them with the baseline, then checks other probes in the same network (or country) and the same probe against other root servers. Fixed rules turn those numbers into a finding that separates observations, hypotheses and unknowns. No AI model is involved, so there are no tokens and no cost.
4. Real telemetry. Each investigation is one OpenTelemetry trace: an agent span and a span per tool call, with timings and errors. It is sent over OTLP to a demo SpanDock server, read back from it, and shown here.
5. Limits, enforced before every call. Investigations per hour, one at a time, tool calls and seconds per investigation. When a limit is reached, the run stops and says so. Refreshing this page never starts an investigation: everyone sees the same shared feed.

Data: RIPE Atlas, public measurements from the RIPE NCC’s probe network. This is a SpanDock experiment using public data; RIPE NCC doesn’t endorse it. Probe positions are shown roughly (to about 10 km), and probe addresses are never shown. The demo server holds only this experiment’s telemetry, never customer data.