Summary
Edith Harbaugh argued that AI-assisted software development should move fast only when guarded by feature flags, kill switches, observability, and human checkpoints. LaunchDarkly’s internal refactoring experiment failed as a one-shot agent task, then succeeded when split into phases, wrapped in safe fallbacks, and supported with better documentation.
Key Topics
5 key topics from Edith Harbaugh at HumanX Amsterdam 2026. Thicker branches were mentioned more often in the talk.
Don't Skip the "Safe" Switch
Click a topic to open it
Key Takeaways
- Break AI coding work into discrete, validated phases.
- Keep humans in the loop until agents earn automation.
- Wrap risky functionality in feature flags for instant fallback.
- Document legacy code so agents have usable context.
- Use agents to reduce safeguard maintenance, not replace engineering judgment.
“The best quote I've heard about AI is it will amplify whatever behaviors you have, good or bad.”
E
Edith Harbaugh
Speaker
Share
Up Next
DEN
WatchNext in agenda
The Ethics Tax: Can You Afford to Build a "Good" Model?
Dylan Fox, Emre Kazim, Nina Müller
ethical aiai governance
FIZ
WatchAlso on agentic coding
When the Barrier to Building Software Falls
Fabian Hedin, Iain Martin, Zivile Einikyte
agentic codingai startups
RRM
WatchRelated
Engineering the Physical World
Rebecca Schaevitz, Rob Pegoraro, Maria Dunford, Fabrizio Pirondini
physical aidigital twins