A cohort difference is a question, not an answer
Cohort tables show that March retained better than June, and that is where they stop. The interesting work is why they differ. The table contains none of the material that would answer that.
Explanation requires everything outside the table: what was shipped, what was priced, which channel customers came from, and what onboarding looked like that month. Those live in four other systems.
Hypotheses are the honest output. Not "March retained better because onboarding improved" but "March coincided with the onboarding change, and the difference is consistent with that", which is a different claim.
Testability is what makes a hypothesis useful. Saying which hypothesis would be confirmed by which measurement turns a story into a plan. Most cohort commentary stops before that.
How the agent is built in Agent Studio
Cohort comparability is the first rule in the system prompt: the minimum size before a difference may be discussed, and the instruction to say "too small to judge" rather than to produce a story about 40 customers.
The system prompt also requires that every hypothesis be labeled with the test that would test it, and that correlational language be explicitly distinguished from causal language.
Prompt skills carry the timeline, releases, pricing changes, channel shifts, and onboarding changes. That is the material the table lacks and the explanation needs.
A flow assembles the material. A PostgreSQL node runs the cohort query. A Math node computes the grid and sizes. A Google Sheets node supplies the timeline. Agent Chat hands all three over.
Notion receives the story with hypotheses, evidence, and test suggestions separated. A Slack node posts only the headline hypothesis.
Durable sessions mean an analyst can push back, "that change shipped mid-month, not at the start," and have the hypotheses revised rather than the whole analysis rerun.
What this agent is built from
- System prompt: sets the minimum cohort size and permits saying a cohort is too small to judge.
- System prompt: requires every hypothesis to be labeled with what would test it, and keeps correlation apart from cause.
- Prompt skills: carry the timeline of releases, pricing, channel, and onboarding changes.
- PostgreSQL (Data): runs the cohort query. Math computes the grid and sizes first.
- Agent Chat (Util): hands the grid, sizes, and timeline over together.
- Notion (Integration): receives hypotheses, evidence, and test suggestions as separate sections.
- When an ActionFlow is enough: If you only need the cohort table, build the workflow. The agent is only for the explanation.
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.