Self-Learning Loop
Definition
Closed-loop feedback mechanism where the SDR agent improves automatically based on downstream CRM outcomes (closed-won / closed-lost) and Gong call recordings — without manual prompt editing by the customer. This is Harmony’s claimed moat against competitors like oneai, elevenlabs, and leaping-ai who push optimization burden onto customers.
Related decisions
- DEC-005 — SDR agent must be self-learning (closed-loop feedback from CRM + Gong)
- DEC-013 — Core market message is “self-learning revenue voice agent”
Sources
- offsite-2026-04-14 — Vitali proposed; team agreed
- vision-2026-04-15 — validated as a real differentiator after leaping-ai deep-dive (their “self-learning” turns out to be manual prompt updates)
Related entities
- vitali — proposed the differentiation thesis
- leaping-ai — claims self-learning, actually does manual updates — Harmony’s foil
- gong — closed-loop signal source
- salesforce — closed-loop outcome source (closed-won / closed-lost)
Related concepts
- self-learning-voice-agent — productized form of the loop
- voice-stack — the proprietary pipeline that lets each conversation stage be tuned independently
- wedge-strategy — this is what makes the SDR wedge defensible
Implementation notes (from infrastructure.md)
- Every playbook + model change versioned for before/after comparison
- Track outcomes by step + segment (objection type, qualification rate, meeting rate, no-show rate, opportunity creation, closed-won/lost)
- Tie call behavior to downstream CRM + Gong outcomes — optimization based on revenue signals, not only call transcripts
- Customer-facing dashboards prove the learning loop works
Evolution
- 2026-04-14: framed as the core differentiator at offsite
- 2026-04-15: validated by competitive analysis — leaping-ai’s false claim sharpened Harmony’s positioning