Skip to main content

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

This diagram synthesises all four preceding maps: outcome defines what we measure, value stream reveals where time dies, dependencies show what blocks fixes, capabilities confirm we can execute.

AGENT & INSTRUMENT DIAGRAM: AGENT PLATFORM
════════════════════════════════════════════════════════════

AGENTS (Yang — who applies force)
─────────────────────────────────────────────────────────
Nav (Dream Agent) ··········· Set priorities, commission outcomes, audit
Engineering Agent (instance) · Build, test, report via drmg CLI
Orchestrator Agent ·········· Route work, resolve blocks, manage plans
VVFL Auditor Agent ·········· Measure enforcement health, extract patterns
Commissioning Agent ········· Walk deployed URLs, verify against PRD
Wik (Human) ················· Architecture decisions, graduation gates, go/no-go

INSTRUMENTS (Yin — what verifies and rewards)
─────────────────────────────────────────────────────────
Priority Table ·········· Build order instrument (top of table = build next)
Session Timer ··········· Recovery time measurement (<30s target, from O1/F1)
Block Signal ············ Broadcast instrument (<2min target, from F2)
VVFL Dashboard ·········· 8-dimension health score (8 auditors, from O4)
Commissioning Table ····· L0-L4 maturity per component (pass/fail per feature)
Outcome Dashboard ······· O1-O4 directly: recovery time, duplication, cycle time, findings
Plan Completion Rate ···· Template effectiveness (% plans that ship vs stall)
Memory Recall Accuracy ·· Did recalled context match what was needed? (from O1)

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

THREE-CHANNEL ARCHITECTURE
┌────────────────────────────────────────────────────┐
│ │
│ Channel 1: FILESYSTEM (spec interface) │
│ ┌──────────┐ ┌──────────┐ │
│ │ Dream │ reads → │ Eng │ │
│ │ Repo │ ← reads │ Repo │ │
│ └──────────┘ └──────────┘ │
│ Priority table = build order │
│ PRD commissioning = pass/fail │
│ │
│ Channel 2: CONVEX MESSAGES (coordination) │
│ ┌──────────┐ 8 types ┌──────────┐ │
│ │ Dream │ ←─────→ │ Eng │ │
│ │ Agents │ status │ Agents │ │
│ └──────────┘ handoff └──────────┘ │
│ blocker │
│ decision │
│ complete │
│ question │
│ context │
│ system │
│ │
│ Channel 3: SUPABASE (measurement interface) │
│ ┌──────────┐ ┌──────────┐ │
│ │ Dream │ reads → │ drmg │ │
│ │ Agents │ │ CLI │ writes │
│ └──────────┘ └──────────┘ │
│ Measurements, plans, patterns, agent state │
│ │
└────────────────────────────────────────────────────┘

AGENT SESSION LIFECYCLE (from value stream)
┌────────────────────────────────────────────────────┐
│ │
│ DISPATCH ──→ BOOTSTRAP ──→ PLAN ──→ │
│ │ │ │ │
│ Priority Session Plan │
│ Table Timer Completion │
│ │ │ │ │
│ ──→ EXECUTE ──→ REPORT ──→ COMMISSION ──→ │
│ │ │ │ │
│ Block Convex Commissioning │
│ Signal Messages Table │
│ │ │ │ │
│ └── FEEDBACK ──────────┘ │
│ │ │
│ Improves next cycle │
└────────────────────────────────────────────────────┘

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

Loop 1: SESSION RECOVERY (per session start)
┌─ Bootstrap → Timer > 30s? → Diagnose memory load → Fix → Bootstrap ─┐
└────────────────────────────────────────────────────────────────────┘
Instrument: Session Timer. Measures: O1 (recovery -80%), F1 (<30s)

Loop 2: COORDINATION (per block event)
┌─ Execute → Blocked? → Signal → Orchestrator routes help → Execute ─┐
└────────────────────────────────────────────────────────────────────┘
Instrument: Block Signal. Measures: O2 (duplicate work -80%), F2 (<2min)

Loop 3: SHIPPING (per feature cycle)
┌─ Plan → Execute → Report → Commission → Gap? → Reprioritise → Plan ─┐
└──────────────────────────────────────────────────────────────────────┘
Instrument: Commissioning Table + Outcome Dashboard. Measures: O3 (cycle time 2x)

Loop 4: LEARNING (weekly/after 3+ runs)
┌─ VVFL Audit → Patterns? → Semantic Memory → Agent Recall → Audit ─┐
│ After 5 runs: zero patterns → improvements? Kill measurement. │
└────────────────────────────────────────────────────────────────────┘
Instrument: VVFL Dashboard + Memory Recall Accuracy. Measures: O4 (50% findings resolved)

Loop 5: GRADUATION (quarterly)
┌─ CLI proven? → API routes → CLI as thin client → A2A wrappers ─┐
│ Kill signal: API doesn't reduce integration friction vs CLI? │
│ → Ship CLI as the product. Don't abstract. │
└────────────────────────────────────────────────────────────────┘
Instrument: Time-to-first-query comparison. Measures: F7 (API parity), F8 (Agent Card), F9 (A2A task)

GRADUATION PATH
─────────────────────────────────────────────────────────

H1: Agent Platform (this PRD)
│ session data, memory patterns, VVFL measurements

H2: BOaaS Operations (same platform, customer-facing)
│ API routes, multi-tenant, Agent Cards

H3: Open Agent Mesh (A2A protocol)
any agent joins via Task Cards, no filesystem needed

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

Agent Roster

AgentTypeForce AppliedVerification
Nav (Dream Agent)AISet priorities, commission outcomes, audit enforcementPriority Table, Commissioning Table, Outcome Dashboard
Engineering AgentAI (instance)Build code, run tests, report status via drmg CLIPlan Completion Rate, Block Signal
Orchestrator AgentAIRoute work to teams, resolve blocks, manage plan lifecycleBlock Signal response time, Plan Completion Rate
VVFL Auditor AgentAIMeasure 8 dimensions, extract patterns, write to memoryVVFL Dashboard, Memory Recall Accuracy
Commissioning AgentAIWalk deployed URLs, capture evidence against PRD featuresCommissioning Table (L0-L4 per component)
Wik (Human)HumanArchitecture decisions, graduation gates, go/no-go, kill signalsAll instruments (final authority)

Instrument Registry

InstrumentMeasuresThresholdAction on FailOutcome Map Link
Priority TableBuild order alignmentTop row = current workReorder if work doesn't match tableGuards all outcomes
Session TimerTime from session start to productiveUnder 30 secondsDiagnose: memory load, message load, or bothO1 (-80% recovery)
Block SignalTime from block to help arrivingUnder 2 minutesEscalate: orchestrator response brokenO2 (-80% duplication), F2
VVFL Dashboard8-dimension enforcement healthAll dimensions measuredIdentify stale/broken: generator, template, rule, skill, agent, platform, virtue, patternO4 (50% findings resolved)
Commissioning TableL0-L4 per component, per PRDL2+ for MVP featuresBlock shipping until commission passesO3 (cycle time 2x)
Outcome DashboardO1-O4 directlyO1≥80%, O2≥80%, O3≥2x, O4≥50%Weekly review: which loop is broken?All outcomes
Plan Completion Rate% plans that ship vs stallAbove 70% completionTemplate audit: wrong template = reworkO3 (cycle time 2x)
Memory Recall AccuracyDid recalled context match need?Above 80% relevantPrune noisy memories, improve relevance scoringO1 (-80% recovery)

Feedback Loop Quality

LoopFrequencySensorActuatorLatency
Session RecoveryPer sessionSession TimerMemory/message loader tuningSeconds
CoordinationPer block eventBlock SignalOrchestrator routingMinutes
ShippingPer feature cycleCommissioning TablePriority table reorderDays
LearningWeekly / 3+ runsVVFL DashboardPattern → semantic memory → recall1 week
GraduationQuarterlyIntegration friction comparisonTransport layer swapQuarter

Activation Sequence

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

PhaseWhat ResolvesAgent StateHuman Covers
Phase 0 (0.5d)--path flag on ETL CLIEngineering agents can load any profileHuman manually starts sessions, routes work
Phase 1 (4.5d)Shared DB, thin router, VVFL auditorsVVFL Auditor online, drmg CLI usableHuman still dispatches, commissions manually
Phase 2 (2d)CLI wrappers, agent moduleEngineering agents use one entry point (drmg)Human reviews plans, approves dispatches
Phase 3 (2d)Priority dispatch, session bootstrapOrchestrator online: dispatch automated, recovery under 30sHuman focuses on architecture and go/no-go
Phase 4 (2d)Pattern extractor, memory writerLearning engine online: cross-session patterns persistHuman reviews patterns, validates learning quality
Phase 5 (2d)Commissioning dispatch, virtue auditorCommissioning Agent online: automated verificationHuman focuses on graduation decisions
Phase 6-7 (5d)API routes, A2A wrappersFull mesh: any agent joins via protocolHuman sets strategy, everything else is platform

Each phase transfers one bottleneck from human to platform. The Capability Map P1 gaps (shared DB, thin router, session bootstrap, priority dispatch, 8-dimension auditors, flexible path, session extract) are the first three phases of transfers.

Gate

Before executing:

  • Every agent named — YES (6 agents: Nav, Engineering, Orchestrator, VVFL Auditor, Commissioning, Human)
  • Every instrument named — YES (8 instruments with thresholds and outcome map links)
  • Feedback loops explicit — YES (5 loops at different frequencies, each with sensor/actuator)
  • Agent-to-agent handoffs documented — YES (three-channel architecture: filesystem, Convex, Supabase)
  • Horizon data flows declared — YES (H1 → H2 → H3: platform → BOaaS → open mesh)
  • Activation sequence links dependency resolution to agent capability — YES (7 phases, each transfers a bottleneck)
  • Each loop instrument links to outcome map success measures — YES (O1-O4, F1, F2 mapped)

Context