From minutes to seconds: Lessons learned optimizing multi-turn agentic workflows

Speed is UX. When a user is forced to wait on an AI response, friction builds and context switching happens. Frustration mounts, mental overhead increases, and even technically impressive tools end up abandoned.

We built a production data-staging pipeline where users can paste in virtually anything. The agent resolves entities, deduplicates entries, links existing records, enriches data via web search, and stages everything in a database-compatible format, waiting for user confirmation or change requests. It worked reliably and handled even the worst inputs gracefully.

Too bad that a single request could take upwards of 3 minutes. Back to the drawing board!

Optimizing agentic AI systems for speed without compromising quality doesn’t just mean choosing a faster model or using the context window more efficiently. At its core, it’s a system design challenge involving state management, interface design, task decomposition, and knowing when not to use an LLM at all.

In this live-demo, you’ll experience a week’s worth of optimization effort condensed into half an hour. In the end, you’ll understand the tools, techniques, mental models, and struggles we experienced along the way.

View the code on GitHub: https://github.com/ResoluteError/weAreDevelopers

Slides: https://docs.google.com/presentation/d/1G6oSbmHhgX3BwG7L3egmf6TPmNizHosUOjtMhsj96JE/edit?usp=sharing