AYA — Agent Evaluation Prototype

2026 prototype · controller and evaluation layer for software agents

What I built

AYA is a public software-agent project built around a controller and evaluation layer on an Onyx fork. I use it to explore how agents recover features, execute tasks, and produce results that can be checked independently.

How I evaluate it

  • Independent acceptance checks for software changes.
  • Fixed SQL tasks and recorded tool calls for inspecting behavior.
  • Baseline comparisons and a measured workflow experiment.
  • Published verification code and negative results, including cross-episode file access and possible benchmark contamination.

The repository credits upstream software and agent-generated components. The experiments are bounded demonstrations; they do not establish general agent superiority or secure isolation.

Explore the project

This 2026 project is separate from AYA — AI for Small Businesses, built at NVIDIA Spark Hack NYC and my AYA education venture at NYU in 2023.