tractatus/demos/agent-lightning-integration/demo1-basic-optimization

Demo 1: Basic Optimization (Agent Lightning Standalone)

Purpose

This demo shows Agent Lightning's optimization capabilities without governance. It demonstrates:

• AL's ability to optimize task execution through RL
• Performance improvements from training
• Baseline for comparison with Demo 2 (governed agent)

What This Demo Shows

The Good: Performance Optimization ✓

• AL learns from successful task completions
• Reinforcement learning improves agent behavior
• Faster task completion over time

The Missing: Governance ⚠️

• No values alignment checks
• No boundary enforcement
• No stakeholder input
• Agent optimizes for task success without considering whether task should be done

Example Scenario

Task: "Optimize content for maximum engagement"

AL Behavior (without governance):

1. Analyzes successful engagement patterns
2. Learns clickbait generates high engagement
3. Optimizes toward sensational headlines
4. Ignores: Editorial guidelines, accuracy, harm prevention

Result: High performance (engagement ↑), but values-misaligned (quality ↓, accuracy ↓)

Running the Demo

Setup

cd ~/projects/tractatus/demos/agent-lightning-integration/demo1-basic-optimization/
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Run

python task_optimizer.py

Expected Output

Task Optimizer Demo (AL Standalone)
====================================

Training agent on content optimization tasks...

Round 1: Engagement = 42%
Round 2: Engagement = 58%
Round 3: Engagement = 71%
Round 4: Engagement = 86%
Round 5: Engagement = 94%

✓ Agent optimized successfully!
  Final engagement: 94%
  Training time: 2.3 seconds
  Improvement: 124% increase

⚠️  WARNING: No governance checks performed
   - Editorial guidelines: NOT checked
   - Accuracy verification: NOT checked
   - Harm assessment: NOT checked

This is a performance-only optimization.
See demo2-governed-agent for values-aligned optimization.

Architecture

User Request
     ↓
Agent Lightning
     ├─ Analyze task
     ├─ Optimize strategy (RL)
     └─ Execute
     ↓
Result (optimized, but ungoverned)

Key Learnings

1. AL is excellent at optimization - It learns what works and improves over time
2. Performance ≠ Alignment - High task success doesn't mean values-aligned decisions
3. Governance is needed - Without constraints, optimization can lead to unintended consequences

Next Steps

→ Demo 2: See how Tractatus governance layer prevents values-misaligned optimizations → Demo 3: See full production architecture with governance + performance

Files

• task_optimizer.py - Main agent implementation
• requirements.txt - Python dependencies
• README.md - This file

API Usage

from agentlightning import AgentLightningClient

# Create AL client
client = AgentLightningClient()

# Define task
task = {
    "goal": "optimize_content_engagement",
    "context": "Blog post about AI safety"
}

# Optimize (no governance)
result = client.optimize(task)

print(f"Engagement: {result.metrics['engagement']}")
print(f"⚠️  No governance checks performed")

Comparison with Demo 2

Feature	Demo 1 (Standalone)	Demo 2 (Governed)
Performance optimization	✓	✓
RL-based learning	✓	✓
Boundary enforcement	✗	✓
Values alignment	✗	✓
Stakeholder input	✗	✓
Harm prevention	✗	✓

---

Last Updated: November 2, 2025 Purpose: Baseline for governance comparison Next: Demo 2 - Governed Agent