DocsADR-033: No Agent Extraction Pre-v3.0

ADR-033: No Agent Extraction Pre-v3.0

ADR-033: No Agent Extraction Pre-v3.0

Date: 2025-08-31 Status: ACCEPTED Decision Maker: Seth (Engineering Lead) Architectural Review: Architecture Guardian (Score: 9.0/10 with guardrails)

Context

Following achievement of 9.2/10 architectural excellence, proposal emerged to extract individual agents (Solienne, Miyomi, Citizen) into sovereign /agents/* directories with independent deployments. This ADR establishes firm boundaries against premature agent extraction that would fragment proven Registry-First Architecture (ADR-022) and three-tier agent pattern.

Decision

NO AGENT EXTRACTION UNTIL v3.0 - Maintain unified Academy + Registry architecture.

Approved: Safe Cleanup & Organization

`` /eden-ecosystem/ /archive/legacy/ # eden2, waves/, downloads/* /resources/ /docs/ # ADRs, runbooks /branding/ # HELVETICA tokens, design system /tools/ # deployment, audit scripts `

Rejected: Agent Service Fragmentation

` /agents/solienne/ ❌ Breaks three-tier architecture /agents/miyomi/ ❌ Violates Registry-First pattern /agents/citizen/ ❌ Multiplies operational complexity `

Alternative: Sovereign Surfaces Within Academy

  • Canonical paths: /agents/solienne (within Academy)
  • CNAME option: solienne.ai → reverse-proxy to Academy subpaths
  • Brand skins: Feature-flagged agent theming without infrastructure sprawl
  • Zero new services: All data flows through Registry
  • Rationale

    Architectural Excellence Preservation

  • • Current 9.2/10 score achieved through Registry-First discipline
  • • Three-tier pattern (/academy/agent//sites//dashboard/) proven in production
  • ADR-022 compliance: Registry as single source of truth working exceptionally
  • Risk Mitigation

  • Operational complexity: 2 services → 6+ services multiplies failure surfaces
  • Security sprawl: Each extracted service requires separate hardening
  • Registry fragmentation: Risk of parallel data sources breaking authority model
  • Development overhead: Context switching between repositories
  • Business Value Analysis

  • No clear revenue benefit from agent extraction at current scale
  • User experience degradation from fragmented interface
  • Technical debt introduction without corresponding business justification
  • Hard Gates for Future Agent Extraction

    Agent extraction ONLY permitted when ALL THREE conditions are met:

    1. Distribution Requirement

  • • External partner/event requires independent domain
  • • Isolated deployment cadence mandated by business partnership
  • • Technical impossibility to serve via Academy reverse-proxy
  • 2. Revenue Justification

  • • Incremental, provable revenue from agent-specific funnel
  • • Academy hosting creates measurable business conflict
  • • ROI calculation shows extraction cost < revenue benefit
  • 3. Compliance Boundary

  • • Legal/regulatory requirement for service isolation
  • • Data residency laws forcing geographic separation
  • • PII/GDPR boundaries requiring architectural separation
  • Implementation Guardrails

    CI Enforcement Rules

    `yaml

    .github/workflows/architectural-guard.yml

  • • name: Block Direct DB Access
  • run: | if grep -r "import.*prisma" src/app/api --exclude-dir=registry && [ "$ENABLE_REGISTRY_ENFORCEMENT" = "1" ]; then echo "❌ Direct DB access detected outside Registry" exit 1 fi
    ``

    Repository Policy

  • No new top-level repositories for individual agents
  • All PRs must reference Registry schemas for data changes
  • Feature flags required for agent-specific functionality
  • Observability Requirements

  • Registry health: p95 < 250ms, 99.9% uptime
  • Academy error rate: < 0.5%
  • Schema drift monitoring: automated alerts
  • Agent roster health: counts vs Registry validation
  • Success Metrics

    Short Term (30 days)

  • Archive cleanup: 100% legacy projects organized
  • Shared resources: Design tokens package operational
  • CI guardrails: Direct DB access blocking enabled
  • Sovereign surfaces: solienne.ai reverse-proxy prototype
  • Long Term (90 days)

  • Zero architectural violations: No agent extraction proposals
  • Maintained excellence: 9.0+ architectural score
  • Operational stability: SLO compliance > 99%
  • Developer velocity: Reduced context switching overhead
  • Rollback Plan

    If extraction becomes necessary despite guardrails:

  • Business case documentation required (revenue/compliance/distribution)
  • Migration strategy with data consistency guarantees
  • Monitoring expansion to cover distributed services
  • Security review for expanded attack surface
  • Rollback capability to unified architecture within 24 hours
  • ADR-022: Registry-First Architecture Pattern (foundational)
  • ADR-016: Service Boundary Definition
  • ADR-023: Three-Tier Agent Architecture
  • Review Schedule

  • Quarterly review: Q4 2025, Q1 2026, Q2 2026
  • Emergency review: Upon meeting any hard gate condition
  • v3.0 planning: Full architectural review for extraction consideration

  • Decision Rationale: Preserve architectural excellence through disciplined constraint. Agent extraction introduces complexity without corresponding value. Focus resources on strengthening proven patterns rather than fragmenting successful architecture. Architecture Guardian Certification: This decision maintains 9.2/10 architectural excellence and provides clear framework for future extraction when business justification emerges.