The Black Box Problem: Making AI Agent Decisions Observable

This talk details an AST-patching framework for capturing execution traces from AI code generation tools, providing essential verification between generation and merging.

Overview

When AI agents write code, you need verification beyond “it compiled.” The tool is an AST-patching instrumentation framework that captures execution traces from AI agents (Claude Code, Cursor, Aider, GPT-OSS, etc.) generating real code changes.

Tech stack