Sequence engines, branching, and human-in-the-loop

Sequence graph primitives

1SEQUENCE GRAPH PRIMITIVES23  Entry: band==A AND valid_email AND not suppressed AND channel_ok4  Step types: email | linkedin_task | call_task | wait | manual_review5  Edges: on_complete → next; on_reply → exit_to_reply_ops;6          on_bounce → suppress; on_unsub → suppress;7          on_no_reply after T → branch_B8  Branches: persona, industry, engaged vs dark, A/B copy arm9  Guards: max_emails_per_domain_per_day; contact.concurrency==11011  Personalization contract:12    required slots filled or step skips to safe generic / HITL13    never send {{empty}} or wrong company name — preflight lint

1. 01Linear — simple cold email; fine for tests, weak for multi-persona.
2. 02Branched — persona/industry/engagement forks.
3. 03Event-driven — product signals or intent re-enter specific arms.
4. 04HITL islands — research step for strategic accounts mid-graph.
5. 05Multi-channel — email + LinkedIn + call with shared stop state.

Personalization at scale

Timing, capacity, and concurrency

1CAPACITY GUARDS23  per_mailbox_daily_send_limit     # warmup + provider caps4  per_domain_daily_send_limit      # your sending root domain policy5  per_target_company_daily_touch   # avoid pile-ons6  per_contact_active_sequences = 17  global_new_enrolls_per_hour      # smooth spikes post-list-load89  Backpressure: if limits hit, queue enrolls — do not drop silently.10  Surface queue depth on the L1 observability census.

Branching that is honest

1. 01Strong signals — reply, bounce, unsub, meeting booked, form fill.
2. 02Medium — click on non-tracking-sensitive assets, inbound site visit matched.
3. 03Weak — open, pixel, time-on-email guesses.
4. 04Research branch — escalate strategic accounts to human before step 3.
5. 05Breakup branch — short exit message, then suppress for cooldown TTL.

Human-in-the-loop design

Copy and offers as system parameters

1. 01Q: How many steps in a cold sequence? Enough to cover angles without normalizing ignore — often 3–5 email touches over 2–4 weeks plus optional multi-channel. Justify with reply curves, not folklore. Stop early on any reply.
2. 02Q: Email vs multi-channel first? Start where identity is strongest and cost is clear. Multi-channel helps strategic ABM but needs shared stop state. Do not duct-tape LinkedIn tasks that continue after email unsub.
3. 03Q: How do you prevent double-sends after integration glitches? Idempotency keys per contact-step, durable sent log, stop flags as source of truth, and reconcile jobs comparing ESP sent events vs engine state.

Exit criteria and cooldowns

1. 01Completed no-reply — cooldown 60–90d before re-enroll; require new intent signal to skip.
2. 02After meeting no-show — short recovery arm owned by AE, not auto cold restart.
3. 03After closed-lost — longer TTL + reason-based messaging if ever re-entered.
4. 04After unsub — permanent email suppress; do not “try a new domain” to dodge.

1ENTRY GUARD (before enroll)23  assert band in {A,B}4  assert valid_email and not suppressed5  assert active_sequences(contact)==06  assert now - last_exited_at >= cooldown_for(exit_reason)7  assert company_daily_touch_count < cap8  assert personalization_preflight_ok(step1)9  else: enqueue_or_reject(reason_code)

A sequence is a state machine that spends reputation. Every step is a withdraw from a finite trust account.