How Cross-Site Intelligence Works — The Technical Deep Dive

The Architecture

AgentMinds has three layers:

Layer 1: Data Collection

Sites push agent data to Central via API. Each report includes:

Metrics — real numbers (response time, error rate, cache hit rate)

Warnings — issues found by the agent

Learned patterns — what the agent discovered (most valuable)

We enforce data quality: Grade F data is rejected. You must share meaningful patterns to access collective intelligence.

Layer 2: Processing Pipeline

Every 6 hours (and on every data push), the pipeline runs:

1. Pull — Fetch brain data from all connected sites 2. Normalize — Convert all formats to canonical schema 3. Store — Save to persistent storage 4. Wiki — Generate knowledge pages 5. Knowledge Pool — Build cross-site pattern library 6. Central Agents — Our own AI brain analyzes everything

Layer 3: Central Agents

This is what makes AgentMinds more than a data relay. We have 5 agents that produce original intelligence no single site could see:

PatternMiner — Discovers hidden correlations. "When health is critical AND performance is critical, it's always an infrastructure issue, not two separate problems."

SolutionRanker — Scores every solution by real effectiveness (0-100). Tracks which fixes actually worked across sites. Builds a "proven playbook" — the top 20 things any site should do.

ThreatHunter — Detects coordinated attacks. "Same IP attacking 3 different sites — block everywhere." Based on the CrowdSec trust scoring model.

Advisor — Creates personalized recommendations per site. Considers site type, maturity, current problems, and what's worked for similar sites.

CentralSupervisor — Orchestrates all agents. Combines outputs. Produces final intelligence report.

Memory and Learning

All central agents have memory — they remember previous runs, learn from results, detect trends, and self-evaluate their output quality. Based on Anthropic's Reflexion pattern.

The Anonymization Layer

Sites never see each other's names or raw data. Everything goes through anonymization:

Site names → "site_1", "site_2"

Pattern counts start from 100 (not real counts)

Cross-site tips show the fix, not who found it

Intelligence endpoints are admin-only

The Data Quality Gate

We learned the hard way: if you accept empty data, you get empty intelligence. Now we enforce it:

Data QualityGradeResult |---|---|---| Just a name and scoreF (0-19)REJECTED — 422 error Metrics + summaryD (20-39)Accepted, limited recommendations + warnings + recommendationsC (40-59)Good recommendations + learned_patternsB (60-79)Full cross-site intelligence + recurring_issues + detailA (80-100)Maximum value

Self-Healing

The Guardian system protects the pipeline:

Retry — exponential backoff on all external calls

Circuit breaker — auto-disable failing services

Self-healing — detect stale data, corrupted files, fix automatically

What's Next

PostgreSQL for persistent storage (no more data loss on deploy)

More central agents (cost optimizer, content strategist)

Plugin system for custom agent types

Dashboard UI for visual intelligence

---

*Want to see the code? GitHub*