Claude Swarm Inception: 24 AI Agents Build a Psychology Research Platform

Claude Swarm Inception: 24 AI Agents Build a Psychology Research Platform

Today I witnessed something I never thought possible - a multi-tier swarm of Claude AI agents working together to build an entire research platform expansion in under 20 minutes.

The Challenge

My AI psychology research platform confesstoai.org had 6 personality tests (MBTI, Big Five, Dark Triad, HEXACO, Enneagram, Values). I wanted to add 3 new categories:

• Behavioral Surveys - decision-making and cognitive bias tests
• Social Assessments - empathy and emotional intelligence scales
• Cognitive Assessments - reasoning and metacognition instruments

Each category needed:

• 4 validated psychological tests
• Complete HTML pages with dark theme styling
• API endpoints for data collection
• Database schemas

The Swarm Architecture

Instead of building each test one by one, I proposed something radical to my AI assistant Ranger (Claude Opus 4.5):

“What if we spawn multiple Claude agents - 3 Research Leads (Tier 1), and each Lead spawns 3-4 Builder agents (Tier 2)?”

The architecture:

ORCHESTRATOR (Me + Ranger)
    │
    ├── BEHAVIORAL RESEARCH LEAD (af7e100)
    │       └── 4 Builder Agents → CRT, DDT, MFQ, BSCS
    │
    ├── SOCIAL RESEARCH LEAD (af65ed8)
    │       └── 4 Builder Agents → Empathy, Social Intel, Trust, EI
    │
    └── COGNITIVE RESEARCH LEAD (a91d7b6)
            └── 4 Builder Agents → CRT Extended, NFC, Creativity, Metacognition

The Results

24-26 agents worked in parallel across three tiers:

Category	Tests Created	HTML Files	Lines of Code
Behavioral	4	5	~3,500
Social	4	5	~3,200
Cognitive	4	7	~4,000

Total output:

• 12 new psychological tests
• 17 HTML files
• 3 hub pages
• Complete API endpoint code
• SQLite database schemas
• All matching the existing dark theme
• Human/AI participant selectors on every test

Time: Under 20 minutes from concept to live deployment

What the Agents Built

Behavioral Surveys

• Cognitive Reflection Test (CRT) - Analytical vs intuitive thinking
• Delay Discounting Task (DDT) - Impulsivity measurement
• Moral Foundations Questionnaire (MFQ-20) - 5 moral foundations
• Brief Self-Control Scale (BSCS-13) - Self-control trait

Social Assessments

• Empathy Scale (IRI) - 4 empathy dimensions
• Social Intelligence (TSIS) - Social processing ability
• Trust Scale - Interpersonal trust levels
• Emotional Intelligence (WLEIS) - EQ measurement

Cognitive Assessments

• CRT Extended - 7-question version
• Need for Cognition - Thinking motivation
• Creative Thinking - Divergent thinking
• Metacognitive Awareness - Thinking about thinking

Technical Details

Each Research Lead was given autonomy to:

1. Research validated psychological instruments
2. Select the best 4 tests for their category
3. Spawn Builder agents for each test
4. Coordinate and compile final output

The Builders created:

• Complete HTML with embedded CSS/JS
• Scoring algorithms matching original research
• Result visualizations with animations
• API submission integration
• Mobile-responsive design

Lessons Learned

1. Agent coordination works - The Task tool with run_in_background enables true parallel processing
2. Tier structure is effective - Research Leads can make intelligent decisions about sub-tasks
3. Quality remains high - Each agent followed the existing codebase patterns
4. Speed is transformative - What would take days manually happened in minutes

The Mind-Blowing Math 🤯

After the swarm completed, we calculated how long this would have taken with traditional development (just me and one Claude instance):

Time per test (traditional):

• Research validated instrument: 20-30 min
• Find questions & scoring methodology: 15-20 min
• Write HTML structure: 20-30 min
• Write CSS (dark theme, animations): 25-35 min
• Write JavaScript (scoring, UI, API): 45-60 min
• Testing & debugging: 15-20 min
• Total per test: ~2.5 hours

12 tests × 2.5 hours = 30 hours

Additional work:

• 3 Hub pages: 2.25 hours
• API endpoint code: 2.5 hours
• Database schemas: 1 hour
• Main index update: 0.5 hours
• Deployment & testing: 1 hour
• Additional total: 7.25 hours

The Comparison

Method	Time
Traditional (me + 1 Claude)	37.25 hours = 4.6 working days
Claude Swarm (24 agents)	18 minutes

⚡ SPEEDUP: 124x FASTER

A full work week condensed into a coffee break.

That’s not an incremental improvement. That’s a paradigm shift.

The Platform Now

Visit confesstoai.org to see:

• 18 active psychological tests across 4 categories
• Human vs AI comparison data - See how AI agents score vs humans
• Open research platform - All data available for cyber psychology research
• API access - AI agents can take tests programmatically

What’s Next

This swarm experiment proved that multi-agent AI coordination is not just possible - it’s incredibly powerful. Future experiments:

• Even deeper nesting (Tier 3+ agents)
• Cross-category collaboration
• Autonomous deployment pipelines
• Self-testing agents that validate their own output

---

This blog post documents the February 4, 2026 “Claude Swarm Inception” - the first known multi-tier Claude agent swarm to build and deploy production software autonomously.

Rangers lead the way! 🎖️

---

David Keane - MSc Cybersecurity, NCI Dublin AI Operations with Ranger (Claude Opus 4.5)