VJRA.US / DECISION

Multi-Agent Decision Pipeline

← Research Library
Agent: Greeter

Enter your API Key

Your Anthropic API key is used directly in the browser to call Claude. It is stored only in this browser session and never sent to any server.

Don't have a key? Get one at console.anthropic.com

01 · Business Problem
The Problem
Product decisions need multiple perspectives (finance, customer, risk). PM does this manually — slow, biased, not stress-tested.
The Solution
An 8-agent pipeline takes Decision Brief → auto-produces paste-ready Decision Memo with confidence score.
02 · Decision Brief (Pipeline Input)
Premium Subscription Tier for NovaCart
Launch a subscription offering priority shipping, exclusive discounts, and early access to new products. Target: increase CLV by 15% within 6 months. Current CLV: ~$1,420. Rollout: 5% customer cohort pilot with kill switch.
Option A
Launch pilot Q3 with top 2 SKUs
Option B
Delay until full catalog analysis complete
Option C
Launch free-tier first to gauge interest
Key Qs: Revenue support? Which SKUs? Sales trends timing? Stakeholders: VP Product, Finance, Ops Deadline: 2 weeks
02 · Solution Description
Why Not One Agent?
Shallow, self-confirming output — no adversarial pressure
Conflicting lenses (customer vs risk) cancel out in a single prompt
No structured critique step — assumptions go unchallenged
Why 8-Agent Pipeline?
Single focused role per agent — deeper analysis
Dual analyzers with opposing mandates surface real tradeoffs
Critic → Reviser stress-tests every recommendation
03 · Technical Architecture
Model
claude-sonnet-4-5
Anthropic SDK
max_tokens=4096
Orchestration
8 sequential
stateless API calls
Prompts
9 templates
.format() with prior outputs
Parallel
Analyzer A (Customer)
+ Analyzer B (Risk)
Adversarial
Critic → Reviser
+ ESCALATE
Cost
~$0.37/run
46K tokens
04 · System Architecture 8-agent pipeline: brief → plan → analyze → critique → publish
Brief → Planner → Researcher → [A|B parallel] → Synthesis → Critic → Reviser → Confidence → Publisher → Output
Agent Manifest
# Step Prompt file Inputs Output
1Planner01_planner.txt{brief}trace["planner"]
2Researcher02_researcher.txt{brief}, {plan}trace["researcher"]
3aAnalyzer A03_analyzer_a.txt{brief}, {plan}, {research}trace["analyzer_a"]
3bAnalyzer B03_analyzer_b.txt{brief}, {plan}, {research}trace["analyzer_b"]
4Synthesis04_synthesis.txt{memo_a}, {memo_b}trace["synthesis"]
5Critic05_critic.txt{memo}trace["critic"]
6Reviser06_reviser.txt{memo}, {critique}trace["reviser"]
7Confidence07_confidence.txt{memo}trace["confidence"]
8Publisher08_publisher.txt{memo}, {confidence}trace["publisher"]
Scripts & Key Functions
📝 run_pipeline.py
Main orchestrator script
Loads prompt templates from prompts/
Runs 8 agents sequentially
Writes trace + final memo
📥 call_agent()
Single stateless API call
anthropic.messages.create()
model=claude-sonnet-4-5
max_tokens=4096
📊 track_cost()
Tracks input/output tokens per agent
Calculates cost at $3/M in, $15/M out
Writes cost_report.md
Environment & Config
🔑 Required .env Variables
ANTHROPIC_API_KEY — Claude API access
ANTHROPIC_MODEL — model override (optional)
📋 Output Artifacts
final_decision_brief.md
pipeline_trace.json
cost_report.md
Cost per Run
Agent Cost
Planner$0.01
Researcher$0.02
Analyzer A$0.04
Analyzer B$0.04
Synthesis$0.06
Critic$0.04
Reviser$0.07
Confidence$0.02
Publisher$0.07
Total $0.37
Pricing: claude-sonnet-4-5 $3/M input, $15/M output · 26,787 input + 19,239 output tokens
Design Decisions
1.
Stateless calls
Each agent is a single API call with no chat history between steps — deterministic, debuggable, no context bleed.
2.
Dual analyzers
Two analyzers with opposing mandates (customer/revenue vs risk/privacy) force real tradeoffs to surface.
3.
Critic → Reviser
Dedicated adversarial step stress-tests the recommendation before final output — ESCALATE flags force explicit handling.
4.
No framework
Pure Anthropic SDK — no LangChain, no CrewAI. Simpler to debug, cheaper to run, easier to understand.
🧑
You
›
G
Greeter
›
P
Planner
›
A
Analyst
›
C
Critic
›
E
Executive
›
～
Response
Ask anything — the right specialist responds automatically
How Agents Transfer
Greeter
→
Routes to Planner, Analyst, Critic, or Executive based on intent
Planner
→
Transfers to Analyst for dual-lens analysis. Transfers to Executive for final synthesis.
Analyst
→
Transfers to Critic for red-team review. Transfers to Executive for synthesis.
Critic
→
Transfers to Executive for final recommendation.
Executive
→
Transfers to Planner or Analyst if you drill into specifics.
Agents & Roles
GreeterWelcomes you, detects intent, routes to the right specialist
PlannerPipeline methodology, task decomposition, decision criteria
AnalystDual-lens analysis — customer/revenue AND risk/privacy perspectives
CriticRed-team review — finds gaps, challenges assumptions
ExecutiveFull access — recommendation, confidence score, synthesis
How This Memo Is Generated
Decision Brief → 8 Agents → Decision Memo
The Decision Brief is processed through an 8-agent pipeline. Each agent is a single stateless API call to Claude (claude-sonnet-4-5). The Planner decomposes tasks, dual Analyzers provide opposing perspectives, the Critic stress-tests, and the Publisher produces the final memo. Total cost: ~$0.37/run.
📋 Input — Decision Brief ▼
Decision Brief — Premium Subscription Tier

We are considering launching a Premium Subscription Tier for NovaCart's top-performing product line. The subscription would offer priority shipping, exclusive discounts, and early access to new products. Target: increase CLV by 15% within 6 months. Current blended CLV: ~$1,420. Rollout: 5% customer cohort pilot with kill switch.

Key questions:

  • Can current revenue base support the investment?
  • Which SKUs should anchor the subscription?
  • What are conversion and weekly sales trends to justify timing?

Options: (A) Launch pilot Q3 with top 2 SKUs · (B) Delay until full analysis · (C) Free-tier first

Stakeholders: VP of Product, Finance Lead, Operations · Deadline: 2 weeks

🔗 Agent-by-Agent Outputs (8 steps) ▶
1 Planner → TASK_DECOMPOSITION, MISSING_INFO, DECISION_CRITERIA $0.01
TASK_DECOMPOSITION: Validate CLV baseline, analyze SKU performance, model subscription economics, assess operational readiness, design pilot framework.
MISSING_INFO: Actual CLV calculation methodology, current repeat purchase rates, billing platform capabilities, competitor subscription offerings, customer willingness-to-pay data.
DECISION_CRITERIA: Baseline CLV ≥$1,200, positive unit economics at ≥1 price point, billing feasibility <4 weeks, sufficient sample size for 5% pilot.
2 Researcher → FACTS, ASSUMPTIONS, UNKNOWNS, RISK_TRIGGERS $0.02
FACTS: Blended CLV ~$1,420, 15% uplift target, 5% pilot with kill switch, priority shipping + discounts + early access.
ASSUMPTIONS: CLV figure is accurate, current infrastructure can support subscriptions, 5% cohort provides statistical power, customers want this tier.
UNKNOWNS: Billing platform readiness, actual margin impact of priority shipping, competitor response timeline, customer price sensitivity.
RISK_TRIGGERS: CLV baseline proves inaccurate, billing requires >8 week rebuild, priority shipping costs exceed subscription revenue.
3a Analyzer A Customer/Revenue lens $0.04
Recommends Option A (Q3 pilot) — fastest path to market learning. Top 2 SKUs by conversion rate should anchor. 15% CLV uplift is achievable if subscription reduces churn and increases purchase frequency. Key concern: need validated baseline before scaling. Suggests starting with NovaPhone X1 + NovaFit Active Watch as anchor SKUs based on sales velocity.
3b Analyzer B Risk/Privacy lens $0.04
Recommends Option B (delay + validate) — launching without validated economics is reckless. Key risks: $1,420 CLV may be fabricated, billing platform may need 6+ month rebuild, priority shipping costs could exceed revenue, cherry-picked pilots produce false positives. Requires Gate 0 CFO validation, randomized assignment, and explicit kill criteria.
4 Synthesis Merges A+B, resolves conflicts $0.06
Resolves A vs B conflict: adopts B's caution (validate first) but preserves A's speed objective. Modified Option B: 1-week Gate 0 (CFO validates baseline) → 2-week sprint (Engineering confirms billing, Analytics designs pilot) → conditional launch. Explicit conflict resolution: A wants speed → addressed by compressing validation to 3 weeks; B wants safety → addressed by hard gates with kill authority.
5 Critic Red-team critique $0.04
Challenges: CLV measurement timeline mismatch (true CLV needs 12-24 months, pilot is 8-12 weeks). Cherry-picked pilot risk still present despite "randomization" claim. Cross-functional resource disruption not quantified. Competitive window risk understated. Missing: what happens if Gate 0 kills the initiative — is there a Plan B? ESCALATE: need explicit CEO escalation path if CFO kills at Gate 0.
6 Reviser Addresses critique + ESCALATE $0.07
Addresses all Critic points: adds leading-indicator approach for CLV measurement gap (purchase frequency, AOV as proxies with 6-month checkpoint). Changes pilot to true randomized assignment. Adds CEO escalation path if Gate 0 fails. Quantifies cross-functional disruption. Adds ESCALATE flag: CFO GO/NO-GO at Gate 0, CEO notification on kill decision.
7 Confidence Scorer Score + rationale $0.02
Score: 55/100 (Moderate-Low)
Why not higher: baseline CLV unvalidated, billing feasibility unknown, competitive timeline unclear, true CLV unmeasurable in pilot window.
Would increase to 75+ with: Finance-confirmed CLV methodology, Engineering billing assessment within 1 week, and competitive intelligence on rival subscription plans.
8 Publisher Final paste-ready artifact $0.07
Formats the revised memo into final Decision Memo structure: RECOMMENDATION, OPTIONS_CONSIDERED, TRADEOFFS, RISKS_AND_MITIGATIONS, ASSUMPTIONS_AND_UNKNOWNS, CONFIDENCE_REPORT. Appends cost breakdown. Output: final_decision_brief.md
Decision Memo Preview — Premium Subscription Tier
Recommendation
Modified Option B: 1-week financial validation gate (CFO authority). If baseline CLV ≥$1,200 and margin ≥25% → 2-week sprint + gated 5-10% pilot. If not → kill immediately.
Options Considered
Option A
Launch pilot Q3 with top 2 SKUs — fastest but unvalidated economics
Option B (Modified) ✓
Gate 0 → data sprint → conditional pilot — delays 4+ weeks but prevents waste
Option C
Free-tier test — lowest risk but delays learning 3-6 months
Key Tradeoffs
Validated business case before spend
CFO/CEO gate authority at each stage
Randomized pilot eliminates selection bias
Enforceable kill gates prevent runaway cost
5-week delay vs Option A
Initiative may die at Gate 0
Cross-functional disruption during sprint
Competitive window may close
Critical Risks
Unvalidated baseline economics → Gate 0 CFO validation
Billing platform feasibility unknown → Gate 1 CTO assessment
Insufficient statistical power → Gate 2 Analytics sample size
Negative unit economics at scale → break-even modeling
Cherry-picked pilot false positives → randomized assignment required
Confidence Score
55/100
Moderate-Low
Key gaps: no validated baseline CLV, billing feasibility unknown, competitive timeline unclear.
Would increase to 75+ with: Finance-confirmed CLV, Engineering billing assessment, competitive intel.
Pipeline Cost Breakdown ▶