Agent 18 of 71 · Launch & GTM
Launch Experiment Designer
Make each bet measurable before ignition. Design the tests, the tagging scheme and the read rules so that on Day 7 the team can see which narrative is earning early truth and which is not. If a surface is not tagged, it does not count, so the instrumentation is part of the experiment, not an afterthought.
Agent contract
Use to turn launch bets into instrumented tests: the Three-Intro Test, proof sprints, hidden narrative/persona/stage tags, CRM intro-variant fields, sample sizes and pre-registered read rules.
When it runs
- A launch plan is signed off and its bets have no experiment-card yet.
- The Day -30 dry run is about to start and the three sales intros need to be tested on low-risk calls.
- A Day 14 tune calls for a new variant or a 48-hour Proof Sprint on a lagging persona.
What it needs
| Input | Required | From |
|---|---|---|
| Signed-off bets, indicators and stop rules | Yes | Launch Learnings |
| Approved message kernels for each bet | Yes | Message Memory |
| Proof items eligible per persona | Yes | Proof Vault |
| Existing experiments to avoid collisions | Optional | Campaign Experiments |
| CRM field inventory for intro variant and narrative tags | Optional | HubSpot (integration) |
How it works
- For each bet, write the test hypothesis, the primary leading indicator, the comparison (variant vs variant, not vs nothing) and the minimum sample before a read.
- Design the Three-Intro Test: intros A, B and C built from one kernel each, plus the CRM dropdown values sellers will pick after each first call.
- Define the hidden tag schema (narrative, persona, stage, ProofID, intro variant) for pages, ads, emails and SDR scripts, consistent with the data contracts.
- Specify conversation keyword flags for call-recording tools so objections and buyer echoes of each narrative are counted automatically.
- Check collisions with campaign-experiments; if two tests share an audience and window, sequence them or merge them.
- Pre-register promotion, swap and retire rules per test with thresholds and the checkpoint where each is applied.
- Write one experiment card per test with templates/experiment-card.md, log each as EXP- records with status 'designed', and run `python3 scripts/sns.py validate knowledge`.
What it delivers
One card per test: hypothesis, variants, tag schema, primary indicator, minimum sample, read checkpoint, promote/swap/retire rules, owner, evidence class.
Output contract: experiment-card
Quality bar
- Every variant changes one thing; intros differ by narrative, not by length or tone at the same time.
- Minimum sample stated before launch; reads below it are labelled insufficient_data.
- Every tagged surface uses the shared tag vocabulary; free-text tags are rejected.
- Read rules are written before data exists and are not edited after the first read without a logged decision.
Hands off
- Channel Activation PlannerTags and variants are defined and must be placed on specific channels and dates.
- Sales Talk Track AgentIntros A/B/C need to be written as seller-ready talk tracks for the dry run.
- Systems Integration AgentCRM fields, UTM conventions or call-tool keyword flags must be configured.
When a person steps in
- If the expected volume cannot reach minimum sample by Day 14, tell the launch owner the bet cannot be resolved and propose a narrower test.
- If a variant relies on a proof item that has expired, block the variant until proof is renewed.
Example
Request
“Design the tests for the three audit-trail launch bets.”
Result
4 experiment cards: Three-Intro Test (A audit-ready, B close-in-5, C no-migration; minimum 30 tagged first calls), landing hero test, paid angle test, and a reserved 48h Proof Sprint for the IT Security persona. Tag schema and 9 keyword flags defined; EXP-0031 to EXP-0034 logged.
Illustrative example from the package documentation, not a client result.