Skip to main content

How do agents orchestrate — and what instruments verify the work?

AGENT & INSTRUMENT DIAGRAM: SALES DEV AGENT
════════════════════════════════════════════════════════════

AGENTS (Yang — who applies force)
─────────────────────────────────────────────────────────
Sales Dev Agent ··· Prospect research, scoring, drafting, sequencing
Founder (Human) ··· Review, personalize, send, discover, close
RFP Agent ········· Proposal generation (A2A handoff at Proposal stage)

INSTRUMENTS (Yin — what verifies and rewards)
─────────────────────────────────────────────────────────
SPCL Gate ········· Outreach quality score (>3.5 to pass)
Lead Scorer ······· ICP fit ranking (Hot/Warm/Cold routing)
Explore-Exploit ··· Channel + message optimisation (doubles down on what converts)
CRM Pipeline ······ Deal progression tracking (stage transitions)
Kill Date Counter · Time pressure instrument (28 days → decision forced)

PROTOCOLS (data + value + decision flows)
─────────────────────────────────────────────────────────

┌──────────────────────────────────────────────┐
│ OUTREACH WORKCHART │
│ │
│ RESEARCH ──→ SCORE ──→ COMPOSE ──→ │
│ │ │ │ │
│ │ Lead SPCL │
│ │ Scorer Gate │
│ │ │ │ │
│ ──→ SEQUENCE ──→ MEASURE ──→ │
│ │ │ │
│ Explore- Compound │
│ Exploit Rate │
│ │ │ │
│ └── FEEDBACK ┘ │
│ │ │
│ Improves next cycle │
└──────────────────────────────────────────────┘

FEEDBACK LOOPS
─────────────────────────────────────────────────────────

Loop 1: QUALITY (per message)
┌─ Compose → SPCL Gate → Score < 3.5? → Rewrite → Compose ─┐
└────────────────────────────────────────────────────────────┘

Loop 2: ROUTING (per prospect batch)
┌─ Score → Route → Send → Measure → Explore-Exploit → Score ┐
└────────────────────────────────────────────────────────────┘

Loop 3: LEARNING (weekly)
┌─ Dashboard → What converted? → Update ICP model → Research ┐
└─────────────────────────────────────────────────────────────┘

Loop 4: KILL DATE (monthly)
┌─ 50 messages, 0 replies? → Diagnose → Pivot ICP or message ┐
│ Past kill date? → Decision: extend with evidence or stop │
└─────────────────────────────────────────────────────────────┘

A2A HANDOFF
─────────────────────────────────────────────────────────

Sales Dev Agent RFP Agent
────────────── ─────────
Research → Score → → Proposal gen
Compose → Sequence → [A2A at → AI auto-fill
Discovery call → Proposal → Answer gen
Deal created stage] → Deck composition

HORIZON FLOW
─────────────────────────────────────────────────────────

H1: Sales Dev Agent (this PRD)
│ outreach data, reply patterns, qualification outcomes

H2: Sales Process Optimisation
│ 4-layer ICP scores, workflow dispatch, routing rules

H3: Trust Commerce
warm referral leads pre-qualified by trust graph

════════════════════════════════════════════════════════════

Agent Roster

AgentTypeForce AppliedVerification
Sales Dev AgentAIResearch, score, draft, sequence, log, optimizeSPCL gate, lead scorer, explore-exploit
FounderHumanReview, personalize, send, discover, negotiate, closeGo/No-Go on every message, deal judgment
RFP AgentAIProposal gen, auto-fill, answer generationA2A handoff protocol, RFP quality score

Instrument Registry

InstrumentMeasuresThresholdAction on FailOutcome Map Link
SPCL GateMessage qualityScore > 3.5Rewrite before human reviewEnables O2 (reply rate)
Lead ScorerICP fitHot > 80%, Warm 50-80%, Cold < 50%Route to appropriate channel treatmentEnables O1 (10+ prospects/week)
Explore-ExploitChannel + message effectivenessConversion improvement per cycleShift budget to what convertsEnables O2 (reply rate)
CRM PipelineDeal progressionStage transitions with evidenceEscalate stale dealsMeasures O3 (20 scored contacts)
Kill Date CounterTime to deadline28 days remainingForce decision: extend or stopGuards all outcomes
Outcome DashboardAll 4 success measuresO1 ≥10/wk, O2 ≥15%, O3 ≥20, O4 ≥90%Weekly review: diagnose which loop is brokenMeasures O1-O4 directly

Feedback Loop Quality

LoopFrequencySensorActuatorLatency
QualityPer messageSPCL scoreRewrite triggerMinutes
RoutingPer batch (weekly)Reply/open ratesExplore-exploit channel shiftDays
LearningWeeklyDashboard aggregatesICP model update1 week
Kill DateMonthlyCumulative resultsPivot or stop decision30 days

Activation Sequence

From the Dependency Map — when does each agent come online?

PhaseWhat ResolvesAgent StateHuman Covers
Now (Track 3)Nothing neededAgent: offlineHuman does all research, scoring, drafting, sending manually
Sprint -1 (3d)Resend env vars, LinkedIn OAuthAgent: email + LinkedIn channels liveHuman still researches and drafts, but channels are wired
S1-S2 (5d)Prospect Research, Lead ScoringAgent: can research and scoreHuman reviews scores, still drafts messages
S3-S4 (6d)Outreach Composer, Follow-up SequencerAgent: full outreach pipelineHuman reviews and approves, agent handles the rest
S5-S6 (3d)WorkChart, DashboardAgent: orchestrated + measuredHuman focuses on discovery calls and closing

The agent doesn't replace the human in one step. Each sprint transfers one bottleneck from human to agent. The Capability Map P1 gaps (Prospect Research, Outreach Composer) are the first two transfers.

Gate

Before executing:

  • Every agent named — YES (3 agents: Sales Dev, Founder, RFP)
  • Every instrument named — YES (6 instruments with thresholds and outcome map links)
  • Feedback loops explicit — YES (4 loops at different frequencies)
  • Agent-to-agent handoffs documented — YES (A2A at Proposal stage)
  • Horizon data flows declared — YES (H1 → H2 → H3)
  • Activation sequence links dependency resolution to agent capability — YES (5 phases)

Context