Week 3 · Day 17/30

Trust & Autonomy Levels

Human-in-the-loop, approval flows, escalation patterns

📅 2026-03-20 ⏱️ 5-6 hodín 📊 Security & Production
Celkový progres 57%

🎯 Cieľ dňa

Navrhnúť trust-based systém kde agenty s vyššou dôveryhodnosťou dostávajú väčšiu autonómiu. Implementovať approval workflows.

core theory

📚 Study Resources

👁️

Galileo — Human-in-the-Loop Agent Oversight

Confidence thresholds, escalation frameworks, architectural patterns pre production.

guide
🦜

LangGraph — Human-in-the-Loop Workflows

Interrupts a Commands v LangGraph: pause, review, approve, resume.

docs
📖

Permit.io — HitL Best Practices & Frameworks

Best practices, frameworks, use cases a demo pre human approval v agentoch.

article
💻

MarkTechPost — Plan-and-Execute with User Approval

Hands-on: LangGraph + Streamlit pre plan-and-execute s explicitným user approval.

tutorial

💡 Key Concepts

Confidence Thresholds — 80-90% range. Financial: 90-95%, customer service: 80-85%, healthcare: 95%+
Autonomy Levels — L0 (manual) → L1 (suggestions) → L2 (auto s approval) → L3 (auto s monitoring) → L4 (full auto)
Interrupts — LangGraph interrupt() function pauzne workflow, prezentuje info humánovi, pokračuje po approval
Escalation Patterns — Confidence-based, risk-based, time-based escalácia. Kedy zapojit človeka.

🔧 Praktické cvičenie

Implementuj approval workflow s LangGraph.

  1. Vytvor agenta s 3 úrovňami autonómie
  2. Low risk actions: auto-execute
  3. Medium risk: notify a pokračuj (async approval)
  4. High risk: pause a čakaj na explicitný approval
  5. Buildni Streamlit UI pre review a approval
  6. Testuj s rôznymi scenármi
AI System Engineer · 30-Day Study Plan · Generated 2026-03-03