Daily curated AI insights you can't miss.
Pure prompts cannot produce stable behavior; eval is needed as a feedback signal
Agent trace logs expose failure causes, preventing new bugs from breaking old features
Small changes to evaluation metrics can completely overturn agent behavior results
This talk describes how agent behavior emerges from a loop of prompts, evals, iteration, and feedback, using the example of a seed-asset agent built by Google.
The agent turns messy advertising creatives — low-quality images, cluttered visuals, heavy text overlays — into clean, reusable assets for downstream generative AI tools.
The speakers share lessons: prompting alone produced unstable behavior; evals were used as feedback signals rather than scorecards; agent trace logs revealed why failures occurred; and iteration was done without breaking fixes already in place.
Speakers include Chris Souza, Preetika Bhateja (product manager), and Daniel Bump (engineer).
This talk describes how agent behavior emerges from a loop of prompts, evals, iteration, and feedback, using the example of a seed-asset agent built by Google.
The agent turns messy advertising creatives — low-quality images, cluttered visuals, heavy text overlays — into clean, reusable assets for downstream generative AI tools.
The speakers share lessons: prompting alone produced unstable behavior; evals were used as feedback signals rather than scorecards; agent trace logs revealed why failures occurred; and iteration was done without breaking fixes already in place.
Speakers include Chris Souza, Preetika Bhateja (product manager), and Daniel Bump (engineer).
