Killing a variant early is the most common way to lose the winner
Optimization rules reduce per-click costs once a fixed number of impressions has been reached. At low volume, that threshold sits well inside the noise, so the rule reliably kills variants before they have a chance to show what they are.
The first decision is whether enough data exists to judge, and rules cannot make that decision. A difference that looks decisive on two hundred impressions is frequently nothing.
The second decision is what the losers have in common. Three variants failing for the same reason is a finding about the audience. Three failing separately is a finding about the copy, and they lead to different next steps.
The third decision is the strategic call: iterate inside the winning hypothesis, or abandon it and test a different argument. Rules only ever do the first.
How the agent is built in Agent Studio
The minimum sample is stated in the system prompt as a hard floor, along with the instruction to answer "not yet" rather than produce a verdict on thin data.
Prompt skills include your hypothesis taxonomy, the same one used when the test set was built, so the agent interprets the results by argument rather than by row.
A Start Scheduled flow assembles the data. An HTTP Request node pulls variant performance data, a Google Sheets node supplies the original hypotheses, and the Agent Chat node passes both to the agent.
The recommendations are sent to Slack. The agent does not pause or reallocate the budget automatically. An agent with spending authority falls into a different risk category than an agent who only has an opinion.
Google Sheets logs every decision with its reasoning, building a record that eventually shows a hypothesis working across campaigns rather than within a single campaign.
Retired variants are kept rather than deleted. An argument that failed in one quarter sometimes works in the next, and only the archive makes that discoverable.
What this agent is built from
- System prompt: states a hard minimum sample and permission to answer "not yet".
- Prompt skills: carry the hypothesis taxonomy used when the test set was built.
- HTTP Request (Util): pulls live variant performance.
- Google Sheets (Integration): supplies the original hypotheses and stores every decision with reasoning.
- Agent Chat (Util): a scheduled flow that hands off performance and hypotheses together.
- Slack (Communication): receives recommendations. No budget moves without a person.
- When an ActionFlow is enough: Rule-based pausing is a workflow that works well at high volume, where the noise problem disappears.
Frequently asked questions
Start building AI workflows
Create a free account, open a template or a blank canvas, and run your first ActionFlow.
Newsletter
Get product updates
New nodes, agents, and product notes. We send mail only when we have something worth opening.