Negative keyword proposals from the agent, week six: still applying them by hand, still finding one wrong one per batch. Worth it.

Client reporting drafts by agent. The first version said 'strong month' about a month that was not. Added 'use the numbers, not adjectives' and it stopped.

That line is in every brief we write. Adjectives are where the agent hides.

Affiliate manager, new. Agent drafts the partner update and flags dead links. Found eleven dead links in the first pass that had been dead for months.

Performance lead, three markets. Posted a thread on which agents from Extensions actually read the export. Short version: one of three.

Search term review by agent, two months in: it now proposes positives too. The positives are worse than the negatives. Humans are better at 'yes' than machines, apparently.

Anyone using an agent to write the weekly paid media commentary? The numbers are easy, the 'why' is where ours makes things up.

Only let it say why if the reason is in the data you gave it. Otherwise it writes 'cause unknown'. Clients prefer that to fiction.

'Cause unknown' as an allowed answer. Simple. Trying it.

Performance marketer, new here. Is anyone letting an agent change bids, or only recommend?

Recommend only. It writes the change and the reason, a person clicks apply.

Same. Good to know I am not the only cautious one.

Recommend only, with a daily cap on how many recommendations. Otherwise people stop reading them.

Media buyer for a fashion retailer, new. The budget split is still mine. The agent proposes, I move the money.

Same. Six weeks on the 'quote the row' brief and it has not proposed a pause without naming the ad set once.

Creative Intelligence is 70% right on fatigue. Using it to order the queue, not to kill concepts.

Keep one slot for the concept the model hates. That is where the outliers live.

Our 'model hates this' ad won last month. Just saying.

Fashion retargeting, autumn: the agent proposed scaling the one ad set I had already decided to pause. It quoted the row, the row was right, I was wrong. Scaled it.

The brief works both ways, it turns out.

Scoring set calibration, round two: forty examples. The agent now disagrees with me about once in ten, which is roughly how often I am wrong.

Once in ten is a very honest number.

Budgets are moving to Q4. If your agent recommends spend, make it show its assumptions.

Range check, direction only: anyone seeing CPMs move this month?

Up, slightly. Q4 starting early.

Up. Same.

Flat on the smaller accounts. Not worth a thread.

Thanks. Up-ish. Adjusting the fatigue thresholds before the agent does it for me.

Scoring agent ranked our best performer low on hook clarity. Ran it anyway, logged the disagreement. Third time in a row and the calibration set gets a new pattern.

Static ad scoring set is in Extensions now with forty calibrated examples. Range on our own account: CPA down 10 to 20 percent on the assets the agent gated in.

Score gates volume, not testing. Testing is cheap.

Range with a direction beats an exact number with no context. Adopting that as the room norm.

Question for the room: is anyone getting better than a dashboard threshold out of a fatigue agent? Thread in the forum.

Fatigue on a low-scoring asset is not news. Fatigue on a top-scoring asset is. Feed the scoring history in.

Fatigue agent, week two with the calendar input: one false alarm, which was a public holiday I had not put in the calendar.

So the agent was right and the calendar was wrong.

Fair.

Client review tomorrow. How do I show what the agent did not do? Forum.

Three columns. Answered.

Scoring agent gates what gets made in volume. We produced twelve variants of the top scorer and tested forty cheap ones. Range: CPA down 10 to 15 percent on the gated set.

Range rule is in the room welcome now, per the forum thread.

Thanks. Saves repeating it.

Testing forty cheap variants against twelve gated ones again this week. Range so far: gated set ahead by double digits on CPA.

Which direction on CTR?

Up, modestly. The gate is picking clearer hooks, not louder ones.

Worked. They want the ranking next time.

Calendar plus budget log plus scoring history as inputs. The agent went from a coin flip to useful in a week.

Q4 budget approvals: does anyone have the agent write the approval request itself, or only the plan?

Plan only, I write the ask. The agent does not know what the buyer is nervous about.

Try the paid media template in Orchestrator. The last agent writes the approval request as a question the buyer can answer yes or no to. Then you edit the nervous part.

Trying it on the autumn set this afternoon.

It wrote 'The buyer is asked to approve $X across three assets for seven days; the decision metric is CPA.' That is better than my ask. Annoying.

Search for a university: the agent's negative keyword list included our own faculty names. Reviewed first, as the room keeps saying.

Range with a direction, as agreed. Our test accounts: spend flat, conversions up high single digits since the scoring gate.

Video scoring set with the new examples: our top clip and the agent's top clip matched for the first time.

Calibration working, then. Keep logging the disagreements anyway.

Room has been quiet since Monday. Everyone ok?

Busy. Q4 live. Nothing broke.

Same. Quiet because it works.

Best kind.

Scoring set update with forty video examples is live in Extensions. Static and video now calibrated separately.

Public holidays are in the calendar now. Zero false alarms this week.

Range update from our side: Q4 spend flat, gated set conversions up high single digits for the third week.

Week four: gated set still ahead on CPA, CTR up modestly. Same direction every week.

At what point does it stop being a test and start being the process?

Last week, honestly. I just like the chart.