Campaigns / AI-led improvement

Test the opening message without giving up control

First email only · agent-prepared · human-approved

An agent can prepare two to four reviewed versions from a campaign outcome. SynthCrew assigns each prospect deterministically, preserves the exact copy on the send, and reports replies or meetings from real provider-accepted traffic.

An experiment never sends by itself

An owner or administrator starts the experiment before scheduling. Campaign readiness, suppression, mailbox capacity, and launch approval still apply. A person also decides whether to apply a measured winner.

What is counted

Denominator: distinct experiment assignments with a real provider acceptance timestamp.
Reply: an inbound reply attributed to the most recent accepted experiment send for that campaign and prospect.
Meeting: a recorded meeting outcome on the attributed inbound conversation.
Not counted: drafts, queued sends, HTTP success alone, synthetic mail, opens, or clicks.

The safe lifecycle

Prepare.Create a setup campaign and reviewed first email. Add one to three challengers; exactly one control must match the current step.
Start.An owner starts the draft in Campaign Studio before any send is scheduled.
Launch normally.The existing human launch checkpoint creates deterministic assignments and immutable scheduled-send provenance.
Measure.Results remain collecting until every version reaches the minimum accepted-send sample.
Decide.A statistically clear recommendation is marked winner ready. An owner can end an inconclusive test without changing the sequence, or apply that exact fresh snapshot.

Let the agent prepare the review

Grant experiments:read for inspection and experiments:write only to an agent responsible for preparing reviewed test definitions. The console remains the human review surface. Mutations also require stable retry headers.

curl --fail-with-body -X POST \
  'https://api.synthcrew.io/api/v2/agent/campaigns/CAMPAIGN_ID/experiments' \
  -H "Authorization: Bearer $SYNTHCREW_AGENT_KEY" \
  -H 'X-API-Version: 2' \
  -H "Idempotency-Key: $SYNTHCREW_IDEMPOTENCY_KEY" \
  -H "X-Agent-Run-Id: $SYNTHCREW_RUN_ID" \
  -H 'Content-Type: application/json' \
  --data @experiment.json

Read the generated request schema from API reference. The control subject and body must exactly match the current first email step; weights use basis points and must total 10000.

Read results

curl --fail-with-body \
  'https://api.synthcrew.io/api/v2/agent/campaign-experiments/EXPERIMENT_ID/results' \
  -H "Authorization: Bearer $SYNTHCREW_AGENT_KEY" \
  -H 'X-API-Version: 2'

collecting: At least one version has not reached the configured accepted-send minimum.
inconclusive: The minimum is met, but the observed leader is not significantly better than every alternative after multiple-comparison correction.
winner_ready: A fresh recommendation exists. It still requires owner review and application in the console.

Stale decisions fail closed

Winner application checks the experiment revision, definition hash, and current observation fingerprint. New evidence makes an older recommendation stale. Refresh results and review again; do not force the old winner.

Current boundary

This release tests the first provider-backed email step only. It does not auto-generate copy, auto-start traffic, rewrite existing scheduled sends, or test follow-up timing. The Agent API can prepare definitions and read evidence, but cannot perform the human start or winner decision.

Open Campaign Studio ↗Agent API guide