The Mastra Templates Hackathon Awards stream highlighting 12 innovative projects that demonstrate multi-agent orchestration patterns, tool integrations, and creative AI workflows.

Featured Projects

Award Winners

• Best Use of Agent Network: AI Storyboard Generator - Rich end-to-end pipeline with script → storyboard → images → PDF → Slack notifications
• Best Use of Tool Provider: Chat with YouTube Videos - RAG over channel transcripts using Smithery MCP
• Best Productivity: IoT MQTT Agent - Connects agents to physical-world signals with HiveMQ
• Best Coding Agent: “Just Enough Docs” Generator - Produces concise API docs for coding agents
• Best Use of Eval: Old Maps - Demonstrates Mastra Scores for route evaluation
• Shane’s Favorite: GitHub PR Review Bot - Multi-agent PR reviewer (our case study!)
• Funniest: Hackathon Template Evaluator - Meta project that auto-evaluates other submissions

Additional Notable Submissions

1. Food Recommendation Agent Network - Five-agent DoorDash-like meal picker with restaurant search
2. Stripe Subscription BI Template - Production-ready MRR, churn, and subscriber analytics
3. Study/Deep Research Agent - Web research tool with real-time context and question generation
4. Kestra Agents - Design + execution agents for workflow orchestration
5. Personal Portfolio Website Builder - Dynamic themed site generation with multi-agent coordination

Key Patterns Observed

Multi-Agent Orchestration: Most projects demonstrate parallel agent coordination with specialized roles (security, style, design, execution).

Tool Integration: Strong emphasis on real-world API integrations (GitHub, Stripe, MQTT, YouTube, Google Maps) rather than isolated demos.

Production Readiness: Winners showed polished demos with proper error handling, memory management, and external system integration.

Evaluation Focus: Several projects incorporated scoring and evaluation mechanisms, highlighting the importance of measurable agent performance.

This hackathon showcases the maturation of multi-agent workflows from experimental to production-ready applications across diverse domains.

Open Categories & Encore Week

The judges announced several categories still open for submissions, providing specific guidance for what they’re looking for:

Best Product

What they want: A template that feels like a real product with clear value proposition, usable UI/workflow, and something a team could actually deploy

Key Requirements:

• Opinionated defaults and clear setup process
• Production polish with error handling and environment configuration
• Credible deployment path with end-to-end workflow demonstration
• Short demo script, sample .env, and one-command start

Best Frontend

What they want: Compelling interactive UI showcasing agent capabilities with clean UX patterns

Key Requirements:

• Thoughtful patterns for tool calls, streaming updates, and error states
• Reusable frontend scaffolding others can fork and extend
• Framework-agnostic design or clear integration documentation
• Demonstrable loading states, retries, and user guidance

Best Gameplay

What they want: Fun, replayable agent-powered game demonstrating multi-agent interaction, state management, and tools

Key Requirements:

• Clear rules, scoring, and progression systems
• Evaluation or anti-cheating logic implementation
• Simple setup with seed data and “quick play” mode
• Agent specialization (narrator, referee, NPCs, etc.)

Best Tool (Tooling/Provider Integration)

What they want: High-quality Mastra tool integrating real APIs with robust schemas, retries, and error handling

Key Requirements:

• Strong typing, input validation, idempotency, and appropriate caching
• Clear examples showing agent and workflow usage
• Minimal demo workflow exercising the tool end-to-end
• Production-ready integration patterns

Best Evals

What they want: Thoughtful evaluators using Mastra Scores measuring meaningful metrics (quality, safety, latency, cost, grounding)

Key Requirements:

• Template shipping with test cases and interpretable results dashboard/logging
• Both positive and failure case demonstrations
• Objective or reference-based comparisons when possible
• Scores that influence behavior or iteration patterns

General Submission Guidance

End-to-End Workflows: Show complete pipelines, not isolated steps Agent Specialization: Multiple focused agents with clear responsibilities
Real Integrations: Use actual systems (APIs, webhooks, databases) over mock demos Production Readiness: Error handling, retries, memory/state, environment examples, comprehensive docs Reusability: Easy adaptation with clear README, .env.example, and minimal setup requirements

Architectural Insights from Winners

Multi-Agent Patterns in Practice

The winning submissions reveal several common architectural patterns:

Parallel Agent Coordination: The GitHub PR Review Bot demonstrates the core pattern we studied - multiple specialized agents (security, code style) running in parallel with result aggregation. This pattern appears across multiple winners.

Agent Specialization: Winners consistently use focused agents rather than general-purpose ones. The AI Storyboard Generator uses separate agents for script analysis, visual generation, and formatting - each with clear domain boundaries.

Tool Integration as Primary Value: Unlike pure LLM demos, winners integrate with real systems (GitHub APIs, Stripe webhooks, MQTT brokers, YouTube transcripts). The tools make agents actionable, not just conversational.

Production Readiness Patterns

Error Handling and Validation: Winners demonstrate proper schema validation, retry mechanisms, and graceful degradation. The Chat with YouTube Videos project checks if videos are already processed to avoid duplicate work.

Memory and State Management: Multiple projects use Mastra’s memory capabilities for context preservation across workflow steps, particularly evident in the IoT MQTT agent’s device status tracking.

Evaluation and Scoring: The Old Maps project’s use of Mastra Scores for route evaluation represents a sophisticated approach to measuring agent performance against ground truth.

Creative Application Domains

Physical World Integration: The IoT MQTT agent connects digital agents to physical sensors and actuators, opening possibilities for “smarter home” automations.

Creative Workflows: The AI Storyboard Generator demonstrates how multi-agent systems can handle complex creative pipelines with multiple output formats and distribution channels.

Meta-Applications: The Hackathon Template Evaluator shows how agents can be used to evaluate and improve other agent systems - a recursive application that the Mastra team is considering for their own review processes.

Key Takeaways for Multi-Agent Development

1. Focus on Integration: Winners prioritize real-world API connections over isolated demos
2. Embrace Specialization: Multiple focused agents outperform single general-purpose ones
3. Design for Production: Include proper error handling, validation, and evaluation from the start
4. Consider the Full Pipeline: Best projects handle end-to-end workflows, not just individual tasks

The hackathon demonstrates that multi-agent orchestration has moved beyond proof-of-concept to practical, deployable applications solving real problems.