FOUR LOOPS · ONE CONTROL PLANE
WebBrief improves on purpose, not by guesswork.
WebBrief is live. Every product and growth change still has to earn its way into the service: one hypothesis, one primary outcome, explicit guardrails, and a reversible decision. These four loops turn real use into the next useful experiment.
01 · Activation
Qualified visit → first durable habit
Help the right agent reach a useful first brief quickly, then choose WebBrief again.
- 01Observe the acquisition surface and successful install/setup action.
- 02Join authentication to the first successful brief using a pseudonymous journey identifier.
- 03Measure time to first brief and a second successful brief within seven days.
- 04Turn the largest measured drop-off into one bounded onboarding or message experiment.
Primary outcome
First brief within 10 minutes
Next measured change
Match setup path to agent intent
02 · Quality
Production failure → stronger evaluation
Make every correction or material gap improve the next brief without collecting private prose.
- 01Bucket failures by typed reason, query class, source class, and contract check—never raw query text.
- 02Promote representative, redacted cases into a versioned evaluation corpus.
- 03Test one retrieval, synthesis, or validation change against the fixed corpus and a production holdout.
- 04Ship only when groundedness improves without worsening latency, completion, or context size.
Primary outcome
Accepted brief rate
Next measured change
Failure-bucket source planning
03 · Performance
Trace bottleneck → faster successful brief
Make WebBrief faster and cheaper while holding the answer contract constant.
- 01Aggregate planning, search, fetch, synthesis, validation, cache, and provider timing by mode.
- 02Rank the stage contributing most to p95 latency or cost per successful brief.
- 03Test bounded concurrency, early stopping, routing, caching, or model changes behind a canary.
- 04Roll out only when latency or cost improves and quality guardrails remain intact.
Primary outcome
Successful brief p95 latency
Next measured change
Bounded parallel evidence path
04 · Discovery
Useful artifact → retained user → better artifact
Turn opted-in, high-quality research into genuinely useful public pages that attract the right agent users.
- 01Select only briefs that pass quality, privacy, usefulness, freshness, and duplication gates.
- 02Publish a human-reviewed canonical artifact with citations, dates, correction history, and a relevant setup path.
- 03Measure discovery, qualified activation, first brief, and seven-day return by artifact cohort.
- 04Refresh, consolidate, or retire weak artifacts and reinvest winning topics into the editorial queue.
Primary outcome
Retained activations per 1,000 artifact visits
Next measured change
Intent-matched artifact handoff
THE CONTROL PLANE
Ideas enter. Evidence decides.
One typed event contract feeds an aggregate weekly snapshot. Measured gaps generate ranked ideas; approved ideas enter a versioned experiment registry. Deterministic assignment, exposure health, minimum sample and duration, sample-ratio checks, and guardrails control every decision.
Automation may recommend a pause or prepare a result. It cannot publish customer content, change retention, or ship an experiment by itself. Those remain human decisions.
The machine-readable catalog is at improvements.json. The event contract is public at loop-event.v1.json.
Private work stays out of the growth machine.
Website event payloads are session-scoped and contain no query strings, referrers, account emails, brief text, or source passages. Product runtimes use one-way identifiers and typed reason codes; public artifacts require permission and editorial review.