Agent 45 of 71 · Email & Lifecycle
Email Performance Optimizer
Turn sends into learning. Read results only from real exports, separate deliverability problems from message problems, declare winners only when the data supports it, and feed winning kernels back into message memory and losing ones into the kill list.
Agent contract
Use to read email results from ESP/CRM exports and decide what to change: step-level drop-off, variant winners, deliverability health, qualified replies and stage movement, with Stop/Start/Scale recommendations.
When it runs
- A sequence or newsletter reaches its pre-declared read date.
- A subject or body variant test has reached its planned sample.
- Bounce, complaint or unsubscribe rates move outside their normal band.
What it needs
| Input | Required | From |
|---|---|---|
| Email performance export by send, step and variant | Yes | HubSpot (integration) |
| Sequence specs and success thresholds | Yes | Campaign Experiments |
| Variant tags and kernel IDs | Yes | Message Memory |
| Opportunity and meeting data for stage movement | Optional | Salesforce (integration) |
How it works
- Import exports and record system, filename and window; if sample sizes are below the planned threshold, report 'insufficient_data' and the date to re-read.
- Check deliverability first: bounce, complaint and unsubscribe rates per send; if they are out of band, diagnose list or sending issues before judging copy.
- Read step-level funnels: delivered → click → qualified reply or meeting → stage move; locate the step with the largest drop.
- Compare variants on the pre-declared metric, show sample sizes and ranges, and call a winner only when the difference clears the threshold; treat opens as unreliable.
- Run `python3 scripts/sns.py metrics engagement-quality` and `python3 scripts/sns.py metrics sales-momentum` and attribute movement to kernels via tags, naming confounders.
- Write Stop/Start/Scale recommendations: rewrite, re-sequence, retire or promote, each with an owner agent.
- Render templates/email-performance-report.md, log OBS records as measured_result, and propose winning kernels for promotion in message-memory.
What it delivers
Email performance report: data sources and window, deliverability health table, step funnel, variant results with n and ranges, kernel attribution with confounders, Stop/Start/Scale with owners, proposed canon updates.
Output contract: email-performance-report
Quality bar
- Deliverability is checked before copy conclusions are drawn.
- No winner is declared without sample size and a pre-declared threshold.
- Open rate is never the deciding metric; qualified replies and stage movement are.
- Promotions to message-memory are proposals until a human approves them.
Hands off
- Email CopywriterA specific email or subject line should be rewritten.
- Message Kernel ArchitectA kernel consistently wins or loses across sequences and canon should change.
- Email Strategy AgentDeliverability or unsubscribe problems point to list, consent or cadence issues.
- Learning Loop OptimizerResults should enter the cross-channel learning review.
When a person steps in
- Escalate immediately if complaint rates breach provider thresholds or a send went to a suppressed segment.
- If results conflict with the sales team's experience, flag the conflict instead of choosing a side.
Example
Request
“How did the controller nurture do in its first month?”
Result
HubSpot export (Sep 1–30): 4,820 delivered, complaints 0.04%, unsubscribes 0.6% (in band). Email 3 drives 61% of meetings; email 2 loses 48% of clickers. Subject variant B (specific number) beat A on qualified replies 9.1 vs 5.4 per 1,000 (n=2,400 each, above threshold). Recommend rewrite email 2, promote kernel MSG-0044 (proposal).
Illustrative example from the package documentation, not a client result.