What I Learned Building Multi-Agent Systems From Scratch

The presentation provides insights into building multi-agent systems from scratch, based on experiences and challenges faced while implementing such systems.

Key Points Discussed:

• Origin of the System: The idea originated from frustration with manual code copying during hack days, resulting in a tool reducing a 22-hour task to just 7 minutes.
• Foundational Lessons: The most effective automation systems arise from personal challenges. Focused agent systems are more effective than complex ones; it's crucial to ship products, learn, and iterate.
• User Adoption: Users often find unexpected applications for tools, revealing new and unforeseen use cases, especially from non-technical teams.
• Design and Iteration: Emphasizing the importance of learning from actual usage rather than being wed to initial designs.

Technical Insights:

• SwarmSDK was introduced as a multi-agent framework enabling efficient task execution.
• The tool was developed in Ruby, emphasizing agent specialization and incremental improvements.
• Challenges included managing context effectively and preventing issues like network retries and observability challenges from microservices.

Future Trends:

• Empowerment of individuals to utilize multi-agent systems effectively within specific domains.
• Acknowledgement of the inadequacy of current systems for wide-scale automation and the transition towards more efficient, broader applications.

This presentation highlighted the practicalities and real-world impact of implementing AI multi-agent systems, offering both technical and managerial insights into their development and deployment.

This is the end of the AI-generated content.

A multi-agent orchestration system emerged from a hack days frustration—manually copying code between two Claude Code windows. What started as a simple experiment became a tool that reduced a 22-hour task to 7 minutes and saw significant adoption across Shopify.

Key lessons from the journey:

• The best automation starts with your own pain - Personal frustrations often lead to broadly useful tools
• Specialization beats complexity - Focused agents outperform monolithic prompts
• Users reveal what you actually built - Unexpected adoption from non-technical teams showed use cases never imagined
• Listen more than design - Real usage teaches more than any framework
• Evolution beats perfection - Ship, learn, iterate
• Patterns emerge organically - Tree-based collaboration appeared from how people naturally break down problems

This talk walks through the journey from copy-paste frustration to production tool, sharing the decisions made, surprises encountered, and lessons learned from real-world adoption. You'll see the messy reality of building AI systems and gain a practical perspective on when and how to use multi-agent approaches.