Meta anomaly watch agent | GTM AI agents

What it does

A monitoring tool that alerts every day is ignored by day four. This agent is built around the opposite contract: if it says nothing, everything is fine, and when it does speak, something is genuinely wrong and worth a decision.

Every morning it compares yesterday against a 14-day baseline that ends the day before yesterday, so a bad day never contaminates its own reference. Volume guards run before any judgment, which is why small accounts and slow days never generate false alarms. When a real anomaly clears the guards, the agent posts one message: the findings ranked by urgency, and at most one funded action.

That action is a proposal, never an execution. Pierce proposes, you approve: the suggested budget change ships with the numbers behind it and waits for a human. When no safe one-tap action exists, the agent says what it would investigate instead of pretending certainty.

Before and after

Before

Dashboards checked when someone remembers to
A broken campaign burns budget for days before anyone notices
Alert tools cry wolf daily and get muted
Fixes depend on whoever is online noticing

With the agent

Every day is compared against a real baseline, automatically
A genuine break surfaces the next morning with the numbers
Silence is the all-clear, so a message actually means something
One proposed, funded action waits for your approval in Slack

How it works

The method is public: it is the same discipline as our open-source meta-anomaly-check skill, run for you every day. Seven gates.

1
Pull yesterday and the baseline
Yesterday as a single day, and the 14 days ending the day before yesterday. A bad day never sits inside its own baseline.
2
Apply volume guards before any judgment
Minimum baseline spend before the account is judged at all, minimum click volume before cost-per-click is judged, and a probability guard before a zero-lead day counts as a signal. Small accounts never page anyone.
3
Verify the pipe before trusting a zero
No data for yesterday is a data problem, not a performance problem. A dead pipeline reads exactly like a dead account, so the agent reports "no data received" rather than a fake zero.
4
Check ranked anomaly classes
Zero leads, then cost-per-lead spike, cost-per-click spike, overspend, delivery drop. Each threshold is a multiple of the baseline, tuned in production, not a gut feeling.
5
Derive at most one funded action
The top anomaly is traced to the campaign that funded it, and the agent works out the one safe move, typically a budget cut on the culprit ad set with a floor it will not cross.
6
Propose, never execute
Pierce proposes, you approve. The action arrives as a Slack proposal with the numbers attached. When no safe one-tap exists, the message downgrades itself to advice.
7
Stay silent otherwise
No all-good messages, no streaks, no summaries. The contract with the reader is that a message means something broke.

The recipe

The open-source method inside this agent:

Terminal

git clone https://github.com/PierreNicolasMarquez/pierce-skills.git
cp -r pierce-skills/skills/meta-anomaly-check ~/.claude/skills/

The detection method is published as a free Claude Code skill: same baselines, same guards, running on your own credentials. Then in Claude Code: "run the meta-anomaly-check skill on my account". The managed agent adds the daily schedule, the funded proposal, and the approval loop.