MMarketingLabs · Agents demo
Live demo · fictional company, real agents

AI agents that analyse, propose and act on paid media — inside guardrails.

Claude-powered agents read 90 days of Meta and Google Ads data for Demo Co, a fictional company with three products and eight planted problems. They find what's wrong, propose budget changes a human approves, and write the weekly report. Every number on this page comes from the data, and every agent is scored.

Loading…

How it works

Each agent does one job and hands off to the next. Nothing touches spend without passing rules written in code, and anything that raises spend waits for a person.

Data
Warehouse
Meta + Google Ads pulled every 2 hours into Postgres, with GA4 and the CRM beside it.
Agent 1
Analyst
Plans, queries with read-only SQL, checks, repeats. Reports the few things that need a decision.
Agent 2
Optimizer
Turns each finding into one sized change — pause, cut, raise — or hands it to the agent that can fix it.
Human
Approval queue
Guardrails trim every proposal. Cuts to losers can apply on their own; anything that raises spend waits for a click.
Agent 3 · in build
Operator
Applies approved changes through the Meta and Google Ads APIs, logged with one-click undo.
Alongside: a Reporter writes the weekly one-page report for a CMO, and a Tracking Watchdog blocks spend increases while pixels or conversions are broken (live accounts only).

Dashboard

What the Analyst sees. Hover the charts for every product's value on a day.

Analyst

Its latest run on Demo Co, scored against the planted problems it can't see.

Loading…

Optimizer → approval queue

The changes it proposed from those findings, after the guardrails, with the numbers each decision rests on. Profit uses each product's value per lead or sale.

Loading…

Weekly report

Written by the Reporter from facts gathered in code. Every number in it is traced back to the data automatically.

Loading…

Guardrails, in code

Rules the model can't talk its way around: they run after it answers.

⚖️
Asymmetric budget rules
Raises up to +20% (+30% for a steady strong performer), with 3 days between raises on Meta and 7 on Google. Cuts up to −50%, deeper only to get back under a cap.
✋
Humans approve spend
Anything that raises spend or turns something on waits for a click. Cuts to anything estimated profitable, or made only to stay under a cap, wait too.
🧾
Monthly caps per product
No proposal can push a product's daily budgets past its monthly cap pace.
📡
Tracking must be healthy
On live accounts, spend increases are blocked while the Watchdog has a failed tracking check.
🔒
Read-only analysis
The Analyst's SQL runs as a SELECT-only database user, one schema, read-only transaction, 8-second timeout.
💵
Cost caps
Per-run cost caps and one monthly AI budget across tools; an agent won't start a run that could break it.