Building Your First AI Agent

Source: YouTube auto-generated captions, lightly cleaned (repeated caption lines removed, misheard brand names corrected).

0:00 — Waste you do not see

This morning, one live account under management showed $905 spent over the last 30 days. $235 went to search terms with zero conversions. 92 distinct terms on negative keyword candidates. Average CPA sits at $65. That quiet waste is the problem an AI agent is built to watch without you living in the interface. A media buyer can find those 92 terms by hand. The issue is cadence. By the time you export search terms, clean the sheet, and add negatives, another week of spend has already moved. Average CPA at $65 stays sticky because the cleanup loop is slower than the auction. An agent shortens that loop without removing your judgment. An AI advertising agent is not

0:48 — Agent versus script

a chatbot bolted onto Google Ads. It is a controlled system that can read account state, propose the next action, and only change something after you confirm. Think of it as a junior buyer who never sleeps, never skips the checklist, and never mutates bids or keywords until you say yes. That boundary is the whole product. Step one is start from a script

1:12 — Step 1 — Start from a script

you already trust. Write the playbook the way you brief a junior buyer. When search terms spend without conversions, draft negatives. When CPA drifts past target, pause the worst ad groups. The script becomes the policy the agent follows. If the policy is vague, the agent will be vague. Clear rules are why this step comes first. Take the live numbers as the first policy test. $235 on zero conversion terms and 92 negative candidates become a concrete rule.

Surface candidates daily, group by theme, and hold for approval. You are not asking the model to invent strategy. You are encoding the strategy you already run, so it executes on time every day. Step two is name every action the agent

2:00 — Step 2 — Name the actions

is allowed to take. In the Google Ads Agent.ai style architecture, that list is finite on purpose. Roughly 28 actions cover read, propose, and mutate paths such as pull search terms, draft negatives, pause keyword, adjust budget cap. A short list is why review stays possible. Unlimited tools become an unreviewable black box. Plain English architecture looks like three layers. Sensors read campaigns, ad groups, terms, and conversion paths. The planner matches what it saw against your script and picks one of the 28 actions.

The actor only runs after confirmation when the action would change the account. Read freely, change carefully.

2:47 — Step 3 — Confirm before mutate

That split is why the system stays safe under real spend. Step three is confirm before mutate on every right. Drafting 92 negative keyword candidates is a read plus proposed path and can be automatic. Adding them, pausing a keyword, or moving budget is a mutate path and must stop for your yes. That gate is why you can let the agent watch overnight without waking up to a changed account you did not approve.

Confirm before mutate also trains trust. Early runs should show you the exact negatives tied to that $235 of zero conversion spend with the CPA context at $65 still visible. You approve or edit once. The agent learns which themes you accept. Over

3:36 — Why this order matters

time, the proposals get tighter because the confirmation log becomes part of the policy. This order matters because each layer bounds the next. Script first stops random tool use. A finite action list stops silent scope creep. Confirm before mutate stops irreversible edits while you are still tuning prompts. If you connect live right access before those three are solid, you are not building an

4:00 — Step 4 — Connect live reads

agent. You are renting uncontrolled automation on top of real spend. Step four is wire the agent to live account reads using the same sensors you already trust in reporting. Pull last 30-day spend, conversion counted terms, and search term waste the way you saw $905 total and $235 of zero conversion spend. The agent should see the same numbers you see in the UI, so proposals map to reality, not to a stale export.

Architecture in plain English stays boring on purpose. A scheduler wakes the agent. Sensors fetch campaigns and search terms. The planner applies your script and selects one allowed action. If the action is mutate, a confirmation card is queued for you. After yes, the actor calls the Ads API once and writes a log line. Boring paths

4:53 — Step 5 — Dry run on waste

are why failures are easy to audit. Step five is run a dry loop on the real waste before you enable rights. Ask the agent to list negative keyword candidates from the last 30 days and rank them by spend. You should see a set that explains most of the $235 with zero conversions. If the list is noisy, fix the script. Do not fix noise by skipping confirmation later. Only after the dry loop matches your judgment, do you open mutate paths one action at a time.

Start with add negatives because that action maps cleanly to the 92 candidates and the wasted $235. Leave bid and budget mutates locked until confirmations on negatives feel routine. Gradual right access is why a first agent does not become a first incident.

5:44 — Step 6 — Daily review loop

Step six is operate with a daily review, not a dashboard addiction. Each morning you get a short queue of proposals tied to live numbers, spend, waste, CPA, and the exact mutate actions waiting. Approve, edit, or reject. The agent continues reading all day. You keep strategy ownership while the 28 actions handle the repetitive motion that used to wait for your next login.

6:09 — Recap

You build the first agent from a script you trust, a finite set of about 28 actions, and confirm before mutate on every right. Connect live reads, dry run on real waste like $235 of zero conversion spend, then open mutate slowly. Full walk-through in the blog is on ahmeego.com. Buddy by Ahmeego, a company of it all started with an idea.