B2B SMB SaaS sales-org-as-agents playbook

Wire a multi-CLI agent stack against a unified messaging API and a LinkedIn scraping API so a small founder-led sales team runs cold outreach, enrichment, lead scoring, and follow-up across WhatsApp, LinkedIn DMs, and Instagram from one harness. The functional split: unipile (single API for WhatsApp + LinkedIn + Instagram DM send), apify (LinkedIn profile and post scraping — most parsers miss post text), and a multi-CLI orchestrator that runs Claude / Codex / Gemini in council mode for the judgment-heavy steps (territory selection, ICP debate, message tone). Per-stage agent roles — research, enrich, score, send, follow-up — each with its own skill and its own message log.

The community-reference playbook (PDF + recording) covers the whole sales funnel for B2B SMB SaaS at AI-native ops cadence. Two reproduction signals from the chat within 48 hours of the recording dropping: one operator connected WhatsApp via unipile and "sent first messages through them"; Instagram DMs also worked. The playbook collapses what would otherwise be SDR + RevOps + tooling sprawl into a single agent harness running on a founder's laptop.

1. Lock the ICP and channels. Pick the two or three messenger surfaces your buyers actually live on (typical SMB SaaS combo: LinkedIn + WhatsApp; sometimes Instagram for prosumer). Don't try to be everywhere at once.
2. Connect unipile as the unified send layer. One API key gets you WhatsApp, LinkedIn, and Instagram DM sending behind one interface. Wire it to your agent harness so messaging code is channel-agnostic — the agent decides which channel per lead, the harness handles the protocol.
3. Connect apify for LinkedIn enrichment. Apify scrapers pull profile fields and post text — most enrichment vendors stop at the profile. Post text is where you find a personalization hook that doesn't read as boilerplate.
4. Split the funnel into per-stage agents. One skill (or sub-agent) per stage: research (find accounts), enrich (pull profile + posts + signals), score (rank against ICP rubric), send (compose + deliver via unipile), follow-up (cadence + reply handling). Each agent has its own ~10-file context catalog — your ICP doc, sample winning messages, the playbook PDF.
5. Wire a council layer for judgment calls. For the steps where one model gets stuck on a single frame — "what's the right territory to open next quarter?", "is this message too aggressive?", "should we DM or wait?" — run Claude + Codex + Gemini in council mode (each answers blind, then a synthesis step). Reuse the existing council plugin pattern; do not invent a parallel one.
6. Cap outreach volume per channel. Especially LinkedIn — premium account upfront, 3–5 new contacts day 1 with delays between, vary messages across leads or you'll trip platform rate limits and burn the account.
7. Log every conversation per lead. One jsonl history file per contact, not one global feed. Otherwise the follow-up agent loses thread on which leads are warm vs cold and over-messages people who already replied.
8. Daily review. The founder still skims the warm-lead queue and reviews top-of-funnel volume before the system runs the next day's batch. Replace the SDR, not the judgment.

Reference video and PDF deck:

AI in Action: Sales Org Autonomy — youtu.be/8BS2ZkrK5Ro
PDF: 20260224_EDU_AISales_OH_ColdOutreach.pdf (shared via Zoom office hours)

LinkedIn outreach rate-limit guidance:

Premium LinkedIn account upfront.
Day 1: 3-5 new contact-initiations only, delays between each.
Vary every message — no two identical bodies.

• Don't auto-send from the scoring agent. Keep send as its own stage; otherwise a misclassification cascades into a bad DM that hits 200 leads before you notice.
• LinkedIn premium is not optional at scale. Free-tier accounts get rate-limited or restricted within days of agent-driven outreach.
• One global message log breaks the follow-up agent. Per-lead jsonl files only. The follow-up agent needs to read "what did we say to this person last?" without scanning unrelated threads.
• Multi-channel ≠ multi-message. Don't pile on WhatsApp + LinkedIn + Instagram to the same prospect simultaneously. Pick one primary channel per lead; fall back to a second only after silence.
• Community MCPs leak keys and break on retry. Same lesson that surfaced in the SEO playbook: for production, have Claude write a thin CLI wrapper over the vendor API. MCP is fine for prototyping; switch to CLI when uptime matters.