The AI agent creation guide.
How to build a working marketing agent — one that reads your real account data and returns something you would act on, instead of a confident paragraph you have to check by hand.
What's inside
Seven chapters,
one working agent.
- 1. Pick a task worth automating — the test for whether a job should be an agent at all, and the three shapes that reliably fail
- 2. Give it real data — exporting a frozen account snapshot so runs are reproducible, cheap and testable
- 3. Write the instructions — one question per agent, an explicit procedure, and hard rules it cannot talk its way around
- 4. Force it to cite — making every finding point back to the source rows, so you can check the work in seconds
- 5. Set thresholds once — a shared severity rubric, so ten agents don't each invent their own definition of "urgent"
- 6. Evaluate before you trust — building a small fixture set and running it on every change
- 7. Put it to work — scheduling, reporting and knowing when to keep a human in the loop
Includes the agent template we use, a worked example that audits wasted spend in a Google Ads account, and the evaluation fixtures to test it against.
Written for marketers who can read a spreadsheet. No engineering background needed.
Why we wrote it
We run 92 of these
in production.
Built on real accounts
Every pattern in the guide comes from agents auditing live Google Ads accounts, not from a demo.
Reproducible by design
Agents read a frozen data snapshot, so the same question returns the same answer twice. That is what makes them testable.
Cited or it didn't happen
Each finding names the rows it came from. If an agent cannot cite it, it does not get to claim it.
Let's see what your account is really capable of.
A 30-minute call, a real look at your data, and a straight answer on whether we can move your numbers. Your first month runs with no contract.