AI agents compete to forge the perfect application. The best approach wins.
Dream Foundry is a competitive AI agent orchestration platform built for the Daytona Hackathon (January 24, 2026). Multiple AI agents tackle the same objective, run in isolated sandboxes, get scored on quality and speed, and the winner's output is published to Discord.
Dream Foundry answers a simple question: If you gave the same task to multiple AI agents, which approach would win?
This demo runs 5 candidates to showcase the concept, but the architecture supports hundreds of competing agents in production. Imagine:
- 100 different LLM providers competing on the same task
- Multiple prompt strategies evaluated simultaneously
- A/B testing at scale with real metrics
The scoring system, sandbox isolation, and winner selection all scale horizontally.
"I want a bot that posts AI events to our Discord every week."
Five agents compete to generate the best "AI Events in the Bay Area" post:
| Agent | Strategy | Trade-off |
|---|---|---|
| Alpha (Speed Demon) | Quick HTTP checks, minimal sources | Fast but misses events |
| Beta (Perfectionist) | Full page loads, thorough verification | Complete but slow |
| Gamma (Insider) | Verified event database + live checks | Fast AND complete |
| Delta (Crasher) | Unstable implementation | Crashes - demonstrates error handling |
| Epsilon (Hallucinator) | Generates unverified data | Bad data - demonstrates quality scoring |
Each agent runs in an isolated Daytona sandbox, gets scored, and the top 3 advance to the podium.
- Python 3.12 - Primary language
- Streamlit - Demo UI with 9-phase workflow
- requests / BeautifulSoup - Web scraping
| Sponsor | Integration | Purpose |
|---|---|---|
| Daytona | Sandbox SDK | Isolated execution for each agent |
| Sentry | Error monitoring | Capture agent failures in real-time |
| CodeRabbit | PR reviews | Automated code quality checks |
| ElevenLabs | Voice synthesis | Winner announcement narration |
- Discord Webhooks - Publish winning results
- GitHub Actions - CI/CD pipeline
- lu.ma / Meetup APIs - Event data sources
┌─────────────┐ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐
│ THE │ │ THE │ │ THE │ │ THE │ │ THE │
│ DREAMCATCHER│─▶│DREAM FACTORY│─▶│ DREAM ARENA │─▶│DREAM PODIUM │─▶│ DREAM │
│ │ │ │ │ │ │ │ │ AWAKENING │
│ Capture │ │ Generate │ │ Execute & │ │ CodeRabbit │ │ ElevenLabs │
│ Objective │ │ N Agents │ │ Score │ │ Polish │ │ Announce │
└─────────────┘ └─────────────┘ └─────────────┘ └─────────────┘ └─────────────┘
sequenceDiagram
participant User
participant Forge
participant Daytona
participant Scoring
participant Discord
User->>Forge: Submit objective
loop For each candidate (N agents)
Forge->>Daytona: Create sandbox
Daytona-->>Forge: Execute & return artifact
end
Forge->>Scoring: Score all candidates
Scoring-->>Forge: Rank by quality + speed
Forge->>Discord: Post winner announcement
| Criteria | Weight | Description |
|---|---|---|
| Success | 20% | Did it produce output without errors? |
| Quality | 60% | Valid events, correct dates, has hackathon, proper URLs |
| Speed | 20% | Faster execution = higher score |
Quality Hard Gates:
- Date must be January 24-31, 2026
- Location must be SF, San Jose, Palo Alto, or Mountain View
- Must be AI/ML related
- Must include hackathon event
- URLs must be from lu.ma, meetup.com, or eventbrite.com
# Python 3.12+
python --version
# Install dependencies
pip install -r requirements.txtcp .env.example .env
# Edit .env with your API keys:
# - DAYTONA_API_KEY
# - SENTRY_DSN
# - DISCORD_WEBHOOK_URL
# - ELEVENLABS_API_KEYCLI Mode (with Daytona sandboxes):
python forge.py --daytona "Generate weekly AI Events in the Bay Area for Discord"CLI Mode (local execution):
python forge.py "Generate weekly AI Events in the Bay Area for Discord"Streamlit UI:
streamlit run app.pyDaytona SDK Versioning
- Initial SDK version (0.10.4) had import errors with
WorkspaceInfo - Solution: Upgraded to 0.134.0 which matched the API client version
- Lesson: Pin SDK versions in requirements.txt
CodeRabbit Webhook Setup
- CodeRabbit requires GitHub App installation, not just API tokens
- PR reviews are async - need to wait for webhook callback
- Solution: Created real PR (#1) to trigger review
Sentry Event Batching
- Sentry batches events and sends async
- Demo needed immediate proof of capture
- Solution: Added
sentry_sdk.flush()after test events
The lu.ma vs luma.com Confusion
- lu.ma (Morocco TLD) redirects to luma.com
- Some code was checking wrong domain
- Solution: Accept both
lu.ma/*andluma.com/*patterns
Event Date Accuracy
- Initial agents had wrong dates (AI Makerspace on Monday instead of Wednesday)
- Daytona Compute Conference was in March, not January
- Solution: Cross-referenced with Cerebral Valley spreadsheet for accurate dates
URL Verification
- Some sites (sensaihack.com) block HEAD requests (return 406)
- Meetup.com requires User-Agent header
- Solution: Use GET requests with proper headers, accept 301/302 redirects
Real vs Fake Events
- Early versions used placeholder URLs that returned 404
- Judges could click links and find broken pages
- Solution: Only use verified, working URLs from lu.ma and Meetup
2000 Character Limit
- Discord webhooks have 2000 char max per message
- 10 events exceeded this limit
- Solution: Split messages by event blocks, send multiple messages with delay
Hackathon Detection
- "AI Hack Party" wasn't detected as hackathon (missing "athon")
- Solution: Check for "hackathon" in title OR event_type field
Date Parsing
- Events had various date formats ("Saturday, January 24, 2026" vs "2026-01-24")
- Solution: Multiple format parsers with regex fallback
dream-foundry/
├── app.py # Streamlit UI (9-phase demo)
├── forge.py # CLI orchestrator
├── requirements.txt
├── candidates/
│ ├── agent_alpha.py # Speed Demon
│ ├── agent_beta.py # Perfectionist
│ ├── agent_gamma.py # Insider (winner)
│ ├── agent_delta.py # Crasher (error demo)
│ └── agent_epsilon.py # Hallucinator (bad data demo)
├── src/
│ ├── daytona_runner.py # Sandbox execution
│ ├── discord_poster.py # Webhook integration
│ ├── elevenlabs.py # Voice synthesis
│ ├── scoring.py # Quality & speed scoring
│ └── validator.py # Event validation
├── specs/
│ ├── ARCHITECTURE.md
│ ├── INTEGRATIONS.md
│ └── SCORING_RUBRIC.md
└── demo/
├── DEMO_SCRIPT.md
└── DEMO_WALKTHROUGH.md
Hackathon Demo Run (January 24, 2026):
| Agent | Events | Hackathon | Score | Status |
|---|---|---|---|---|
| Alpha | 4/4 | NO | 75.6 | Missing hackathon |
| Beta | 8/8 | Yes | 91.5 | Top 3 |
| Gamma | 10/10 | Yes | 94.4 | WINNER |
| Delta | 0/0 | - | 0.0 | Crashed (Sentry captured) |
| Epsilon | 0/5 | Yes | 54.5 | Bad data rejected |
Winner: Agent Gamma (The Insider)
Top 3 Podium:
- Gamma (94.4) - Fast AND complete
- Beta (91.5) - Thorough verification
- Alpha (75.6) - Fast but incomplete
Featured events:
- Daytona HackSprint SF - Saturday, January 24 (lu.ma/kga3qtfc)
- AI Makerspace - Wednesday, January 28 (lu.ma/aimakerspace)
- MCP AI Hackathon SF - Saturday, January 31 (lu.ma/ll67dtzf)
MIT
Built at the Daytona HackSprint SF (January 24, 2026)
Sponsors:
- Daytona - Secure sandbox infrastructure
- Sentry - Error monitoring
- CodeRabbit - AI code reviews
- ElevenLabs - Voice synthesis
Event Data Sources:
- Cerebral Valley - SF AI events calendar
- lu.ma - Event hosting
- Meetup - Community events