Economics Framework
Methodology для принятия sizing decisions (agent vs human, build vs buy, expand vs pause). Не hardcoded numbers — framework для применения к Synth Nova sizing questions.
Framework источник: Reference-Org-Blueprint §10.
Context
По мере роста Synth Nova появляются sizing questions:
- Автоматизировать новую функцию через agent или оставить manual?
- Нанимать ли contractor для narrow task или делегировать existing agent?
- Build own tool или использовать SaaS?
- Scale up existing agent tier или создать новый?
Answer depends на unit economics + Manifesto принципы. Framework makes decisions structured, не gut-feel.
Unit Economics Components
Cost components
Per-task (variable):
- LLM inference cost (tokens × model tier)
- External tool costs (API calls, rate-limited services)
- Human-in-loop cost (Founder / Director time)
Per-agent (semi-fixed):
- Agent manifest development / maintenance
- Prompt tuning cycles
- Observability overhead (tracing, logging)
Per-infrastructure (fixed):
- Event bus operations
- RAG / vector DB
- Trace store
- Hosting / compute
Output measures
- Task success rate — actual / attempted (target ≥85% per Manifesto)
- Time to validation — trigger → first useful data (target ≤72h)
- Human intervention rate — % tasks requiring HITL (target <20% через 6 мес)
- Cost trend QoQ — should decline per Manifesto
- Zero security incidents — Codex compliance
Decision Questions
Sequentially. If any answer is “no” — не automate (или не expand).
Q1: Does this task recur frequently enough?
Per Manifesto — “не автоматизируем разовое”.
Rule of thumb:
- Weekly+ recurrence → candidate for agent
- Monthly 1-3 times → borderline, consider cost
- Quarterly or less → probably manual / contractor
Q2: Is this reversible?
Per Rules-Criticality и Rules-AgentDecisionBoundaries:
- Reversible → agent OK (L1-L3 range)
- Hard to reverse → HITL mandatory (L3+)
- Irreversible → human required (L4+)
Irreversible не делаются агентами даже если ROI высокий. Cost of single mistake может wipe out all savings.
Q3: Can we achieve ≥85% success rate?
Piloting required. Before committing к agent:
- Shadow mode (agent proposes, human acts) — measure predicted quality
- A/B if possible
- If <85% attainable → or defer, or лучший model tier, or hybrid (agent + human-in-loop)
Don’t deploy агента с reliability под target. It erodes trust в system generally.
Q4: Is agent cost < human cost (fully loaded)?
Human fully-loaded: salary × benefits × overhead (~1.3-1.5× base comp в US/UAE) + opportunity cost (founder time ≫ hourly rate).
Agent fully-loaded: LLM inference + infra amortized + development/maintenance + HITL fraction cost.
Simple check: if task takes human X hours × rate, agent должен cost < X × rate.
Caveat: human cost for Founder task ≠ hourly — это opportunity cost (что ещё мог бы делать). Для Founder-level decisions automation рate-of-return часто заоблачный.
Q5: Does automation preserve optionality?
Agents — investment. But:
- Locks-in architecture (future changes cost)
- Creates dependency (if LLM pricing shifts)
- May ossify process (“automation tax” to change)
Rule: first automate stable processes. Churning processes keep manual until stable.
Decision Matrix
Сводим Q1-Q5 в матрицу:
| Recurrence | Reversibility | Success ≥85% | Cost favorable | Process stable | Decision |
|---|---|---|---|---|---|
| High | Reversible | Yes | Yes | Yes | Automate (L1-L2 agent) |
| High | Hard-reverse | Yes | Yes | Yes | Automate with HITL (L3) |
| High | Irreversible | Yes | Yes | Yes | Propose-only (L4, human decides) |
| Low | Any | Any | Any | Any | Manual / contractor |
| Any | Any | No | Any | Any | Defer / improve pilot |
| Any | Any | Yes | No | Any | Manual (ROI negative) |
| Any | Any | Yes | Yes | No | Stabilize first, then automate |
Applied per candidate automation. Не one-off decision — review quarterly.
Examples of Application
[illustrative] — показать framework mechanics, не prescribe decisions.
Example 1: Automate competitor research
- Q1 recurrence: Weekly per active niche → high ✓
- Q2 reversibility: Research output — reversible (can re-do) → reversible ✓
- Q3 success ≥85%: Pilot showed 92% quality (verified by Agent-Judge) → yes ✓
- Q4 cost: Agent ~10/wk vs contractor $200/wk → agent wins ✓
- Q5 stable: Research process mature → yes ✓
Decision: Automate (L1-L2 agent). → Agent-MarketResearcher.
Example 2: New niche entry decision
- Q1 recurrence: One-off per niche → low ✗
- Q2 reversibility: Hard to reverse (committed resources, positioning) → hard-reverse
- Q3 success ≥85%: Can’t predict (novel situation) → uncertain
- Q4 cost: Founder judgment high value here → agent inferior
- Q5 stable: Novel → no
Decision: Manual — Founder judgment + Chamber advisory. Agents прapare materials и research only.
Example 3: Automated social media posting
- Q1 recurrence: Daily → high ✓
- Q2 reversibility: Published content — hard to fully unpublish reputation → hard-reverse
- Q3 success: Need quality bar — brand voice, factual accuracy
- Q4 cost: Agent < human, but public visibility raises stakes
- Q5 stable: Brand voice evolving → no (yet)
Decision: Defer automation. Manual или agent-drafted + human-approved (L3 HITL) интерим.
Portfolio view
Don’t optimize per-task — optimize portfolio.
- Fixed infrastructure cost — amortizes across всех automated tasks
- Learning — каждый automated task improves system (calibration, patterns)
- Compounding — automated tasks free Founder time for strategic work
Portfolio decisions:
- Next automation candidate = highest ROI tree-climbing
- Don’t automate everything — retained manual gives flexibility
- Periodic “automation audit” — some decisions reverse if cost shifts
Integration с Manifesto metrics
Framework decisions directly tied к Manifesto success metrics:
| Metric | Framework input |
|---|---|
| Task success rate ≥85% | Q3 — must achieve before automating |
| Cost per task declining QoQ | Q4 — framework forces cost discipline |
| Time to validation ≤72h | Q1 — frequent tasks automated = fast |
| Security incidents zero | Q2 — irreversible stays human |
| HITL rate <20% | Q3 + Q5 — high success + stable process = less HITL |
Decisions misaligning framework с metrics = red flag.
[illustrative] numbers caveat
Blueprint origin содержит classic vs AI-first comparison для hypothetical $50M ARR B2B SaaS. Key magnitudes:
- Classic opex ~36M, delta ~$41M/yr
- CAC 3x reduction
- Revenue/FTE 3x increase
Эти numbers не target Synth Nova. Vertical и phase discrepant. Use их как illustration framework mechanics, не benchmark.
Synth Nova specific numbers calculated on-demand через применение framework к actual questions. Document результаты в Decision-Log или ADR где material.
Open Questions
- Fully-loaded agent cost accounting — how разнести infra cost per-task? (Allocation method нужен)
- Opportunity cost Founder time — дolженlarized dollar amount для comparison?
- Long-horizon ROI — как value agent который learns over time (initial unit economics poor, но compounding?)
- Multi-niche scaling — framework per-niche или shared?
- Contractor vs agent decision matrix — specific middle ground
Связанные документы
- Manifesto — metrics targets
- Reference-Org-Blueprint — section §10 source
- Agent-vs-Human-Tradeoffs — complementary decision matrix
- Rules-Criticality — reversibility classification
- Rules-AgentDecisionBoundaries — what agent can do
- Rules-Budget — budget constraints
- Process-OutcomeLabeling — feeds framework inputs (actual success rate)
- Observability — metrics collection
- Build-Measure-Learn
- Lean-Startup
- Hypothesis-Driven-Development
- Synth-Nova-Overview
- MVP-Phase
- ADR-0020-reference-architecture-blueprint