A complete audit trail has one operational weakness: completeness. At production volume, nobody reads the tape. So Wytness reads it for you — a background engine that scans every active agent's recent behaviour against seven rules and raises its hand when something deserves a human. Here are the seven, in plain language.
The seven rules
- 1 · New tool. An agent used a tool it hasn't touched in the last thirty days. Usually a deploy; occasionally the first visible sign of an agent doing something nobody asked it to. Low severity — a nudge, not an alarm.
- 2 · Frequency spike. An agent's activity this hour is far above its own seven-day average. The baseline is per-agent — a busy agent being busy is normal; a quiet one suddenly running hot is not. Runaway loops, retry storms, and misconfigured schedules all announce themselves here first.
- 3 · High error rate. Too many failures in the agent's last hundred actions. Broken integrations and confused agents both look like this, and both deserve attention before customers provide the alert for you.
- 4 · Restricted data. An event touched inputs classified as restricted. High severity, and the alert links straight to the event in question, so the first click of the investigation is already done.
- 5 · Off-hours activity. An agent acting at an hour that historically accounts for almost none of its activity — with a minimum-history guard, so a new agent's first week doesn't trigger false 3am stories.
- 6 · Chain break. An event's hash doesn't link back to its chain. This is the tamper-evidence alarm, and the only rule rated critical: it means the record itself — not the agent — needs attention, now.
- 7 · Capability deviation. An agent's recent actions fall outside the behavioural profile built from its own history. The widest net of the seven: it catches "this agent is not acting like itself" even when no single action looks wrong.
Why baselines beat thresholds
Notice what most of the rules have in common: the comparison is the agent against itself. Fixed thresholds produce alert fatigue, because a fleet of agents has no single "normal" — a batch processor's healthy Tuesday would trip any limit tuned for a support bot, and a limit loose enough for both catches nothing. An agent's own thirty-day history does the tuning for you. The engine's job isn't to know what bad looks like in general — it's to notice when an agent stops looking like its own past, and to say so with the evidence attached.
Fatigue is designed against in the plumbing, too: alerts deduplicate per agent and rule, so a condition that persists shows up as one item to deal with, not a page of copies of itself. A quiet queue is the goal — when everything is normal, the page should say so by being boring.
Every alert is a shortcut, not a verdict
An anomaly here is never "we've decided this is an attack." It's "a human should look, and here's exactly where." Each alert names the agent, the rule, the severity, and links to the events behind it — which, because this is Wytness, are signed and chain-checked records rather than log lines. The triage rhythm that falls out is pleasantly mundane: skim the queue with morning coffee, critical first; most items resolve in a click — a deploy explains the new tool, a product launch explains the spike — and the one that doesn't resolve in a click is the one the queue exists for. The engine finds the needle; the tape proves it's real.
And because the alerts live in the same system as the evidence, the anomaly history becomes evidence itself: your Evidence Packs carry the period's alerts alongside their resolutions. "We monitor agent behaviour" is a policy claim; a queue of dated alerts, each with its outcome, is what monitoring looks like when it can prove it happened.
Questions about tuning the rules to your fleet? Ask us.