Code Generation Agents: Architecture, Data Modeling Challenges, and Production-Ready Considerations

This session provides an in-depth look at the architecture of multi-agent systems for code generation, emphasizing practical solutions to common data modeling challenges. We’ll explore how focusing on signature data, employing linters, indexing codebases, and utilizing GraphRAG can address issues often faced in agent-driven coding environments. We’ll examine examples where models fall short, and show new concepts that resolve these challenges. You’ll leave with a practical understanding of how multi-agent systems are structured, as well as insights on minimizing bugs and enhancing reliability.

Lee Twito, GenAI Lead at Lemonade, and Alon Gubkin, Co-Founder & CTO at Aporia, discuss the practical aspects of building and deploying code generation agents. The talk covers the retrieval process, including building a codebase graph, identifying relevant files, chunking code, and enriching it with contextual information. It also delves into the coding and testing phases, discussing model selection, edit formats, and various error types. Finally, the speakers touch upon production readiness and future directions for fully autonomous AI software engineering.