Two-Tier Intelligence Architecture

Why IFP works differently than traditional monitoring.

The Core Claim

IFP uses two-tier agent-native intelligence architecture.

Most monitoring tools have single-layer detection systems.

IFP combines strategic meta-observer with tactical specialists.

SAGE observes infrastructure AND the agents themselves.

Agents execute in domains with learned confidence weights.

This creates recursive meta-observation capability unavailable elsewhere.

Result: Predicts problems before they cascade across systems.

Tier 1: SAGE (Strategic Meta-Observer)

Property Value Verification
Scope Infrastructure-wide (everything) Live dashboard
Frequency Every 60 seconds (1,440 cycles/day) Process ID: 335021
Depth Recursive reasoning (1-10+ levels) Technical details
Automation 0% (suggests only, never executes) logs/daemon_output.log
Intelligence Claude Sonnet 4 (deep reasoning) API usage logs
Learning Context accumulation (25+ days) insights.jsonl (25,000+ entries)
Production Status 12,960 cycles completed 9 days continuous operation

Example SAGE Decision

Observation:
├─ DevOps Agent restart success: 87% → 79% (3-day trend)
├─ Correlation Agent: Pattern discovery CPU spikes
└─ TimescaleDB: Feature extraction every 5 minutes

SAGE Analysis (Depth 3):
"DevOps Agent success rate decline correlates with
 feature extraction cycles. ML jobs cause temporary
 CPU spikes, increasing container restart failures.
 This is not declining agent performance - it's
 environmental correlation."

SAGE Suggestion:
"Increase DevOps Agent confidence threshold from
 0.80 to 0.85 during ML training cycles. Schedule
 feature extraction during low-activity periods."

Result: Success rate recovered to 89% (better than baseline)

Tier 2: Agent Layer (Tactical Specialists)

Agent Domain Success Rate Observations/Day Status
DevOps Agent Docker containers (35 monitored) 87% 1,440 ✅ Operational
Web3 Agent Blockchain wallet monitoring 95% 2,880 ✅ Operational
Correlation Agent Cross-domain pattern detection 73% 720 ✅ Operational
Total 3 domains + meta 85% avg 4,680 30+ days runtime

Three-Loop Reasoning Pattern

LOOP 1: OBSERVE
├─ Collect domain telemetry
├─ Frequency: 30-120 seconds per agent
└─ Output: Structured telemetry data

LOOP 2: ANALYZE
├─ Detect patterns in telemetry
├─ Generate insights with confidence scores
└─ Output: List of actionable insights

LOOP 3: ACT
├─ 3a. SUGGEST: Propose remediation actions
├─ 3b. EXECUTE: Take action if confidence high
├─ 3c. LEARN: Update weights from outcomes
└─ Output: Action results + learning

Why Two Tiers Matter

Traditional Monitoring (Single-Tier)

Alert: High CPU on server-5
Human investigates
Realizes: Scheduled backup job
Manual action: Adjust backup schedule
Time: 2-4 hours (human involvement)
Learning: None (forgotten next time)

IFP Two-Tier Intelligence

DevOps Agent: High CPU on ifp-feature-extraction
SAGE observes: This happens every 5 minutes
SAGE reflects: Pattern discovery ML job
SAGE correlates: Web3 transaction delays coincide
SAGE suggests: Schedule ML jobs differently
Human approves: One-time decision
SAGE learns: Stores pattern for future
Time: 5 minutes (first occurrence)
Learning: Permanent (automated thereafter)

Production Verification

📊 Live Metrics

Real-time SAGE dashboard shows current state.

View SAGE Dashboard →

📁 Source Code

Full implementation available on GitHub.

View on GitHub →

📈 Agent Performance

Live agent success rates and observations.

View Agent Metrics →

📝 Decision Logs

Context accumulation: insights.jsonl (25+ days)

~/ifp-workspace/insights.jsonl