AI DevSummit 2025 + DeveloperWeek Leadership 2025

YK Sugi, CS Dojo, Founder

Building an autonomous software engineering agent has been full of learnings. In this talk, I’ll share five unexpected lessons we’ve learned while developing an agent that takes actions on behalf of users.

First, using LLMs effectively requires a mindset shift: rather than just a cost, it’s an investment that needs leadership support and a culture that encourages experimentation.

Second, they’re surprisingly good at generating mermaid diagrams, even mapping out entire repositories.

Third, they expand the scope of what’s practical to build—making it easier to create things that would have been too time-consuming before.

Fourth, when you give people a capable agent, they'll build amazing things with it, even improving the agent itself.

And finally, using an agent to work with large codebases is possible, with a few caveats—test-driven development helps, and software engineering skills still matter.

I’ll wrap up with a live demo: fixing the agent using itself.