Skip to content

AGENTS_GUIDE.md

Latest commit

 

History

History
85 lines (66 loc) · 3.49 KB

AGENTS_GUIDE.md

File metadata and controls

85 lines (66 loc) · 3.49 KB
Raw

🤖 LoomiFlow AI — Agents Guide

Loomi Connect AI Hackathon 2026

This guide details the internal logic and responsibilities of the 4 intelligent agents forming the core of the Autonomous Commerce Operations Agent (ACOA).

1. Agent Roles & Logic

🛡️ Fraud Agent (Weight: 62%)

  • Input: PayPal signals, velocity metrics, IP/Geo data, device fingerprinting.
  • Logic: Strict risk evaluation.
    • score > 0.8 → BLOCK
    • score > 0.4 → STEP_UP_AUTH
    • score <= 0.4 → ALLOW

💰 Revenue Agent (Weight: 23%)

  • Input: Customer LTV, tier (VIP, standard), cart value, transaction history.
  • Logic: Maximizes recovery and safeguards VIP accounts.
    • LTV > 1000 or value > 200 → ALLOW
    • Else → HOLD

💬 CX Agent (Weight: 15%)

  • Input: Churn risk prediction, engagement score, recent support tickets.
  • Logic: Minimizes friction.
    • churnRisk === "high" → STEP_UP_AUTH (to provide a soft-fail path)
    • Else → HOLD

2. Orchestrator Consensus & Matrix

The Orchestrator resolves conflicting advice by applying predefined safety and business rules.

Decision Matrix (Fraud Score vs. LTV)

                    FRAUD SCORE
              0.0    0.3    0.5    0.7    0.85   1.0
           ┌──────┬──────┬──────┬──────┬──────┬──────┐
    < 500  │ HOLD │ HOLD │ HOLD │ BLOCK│ BLOCK│ BLOCK│
           ├──────┼──────┼──────┼──────┼──────┼──────┤
L    500  │ALLOW │ALLOW │ HOLD │ STEP │ STEP │ BLOCK│
T   1000  │ALLOW │ALLOW │ HOLD │ STEP │ STEP │ BLOCK│
V         ├──────┼──────┼──────┼──────┼──────┼──────┤
  > 1000  │ALLOW │ALLOW │ALLOW │ STEP │ STEP │ STEP │
           └──────┴──────┴──────┴──────┴──────┴──────┘
          STEP = STEP_UP_AUTH (VIP protection pattern)

Confidence Calculation

confidence = (0.62 * fraud.conf) + (0.23 * revenue.conf) + (0.15 * cx.conf)

3. The "WHY" Button (Reasoning Chain)

The UI features a "WHY" button powered by the Memory Graph. It calculates the relative influence of every context data point on the final decision.

Example VIP Output:

"Fraud score 0.72 exceeds threshold but customer LTV (€3200) justifies recovery. CX Agent highlighted high churn risk. Final decision: STEP_UP_AUTH."

4. Console Debugging (Expected Prints)

When the agents run, expect the following server logs:

[AGENT][FRAUD][START] {"eventType":"payment_failed","customerId":"vip_pacific_001"}
[AGENT][FRAUD][DONE] {"score":0.72,"recommendation":"STEP_UP_AUTH"}
[AGENT][REVENUE][START] {"eventType":"payment_failed","value":249.9}
[AGENT][REVENUE][DONE] {"revenueAtRisk":249.9,"recommendation":"ALLOW"}
[AGENT][CX][START] {"churnRisk":"high"}
[AGENT][CX][DONE] {"churnRisk":"high","recommendation":"STEP_UP_AUTH"}
[ORCHESTRATOR][START] {"fraud":"STEP_UP_AUTH","revenue":"ALLOW","cx":"STEP_UP_AUTH"}
[ORCHESTRATOR][DECISION] {"final":"HOLD","confidence":0.89,"severity":"medium"}

📝 Proof of Capture

  • Agent specific logic, weights, and inputs.
  • Orchestrator decision matrix.
  • Confidence consensus calculation.
  • Reasoning chain explanation.
  • Expected debug print statements.