Overview
Agent Foundry — Debug your AI agents like software.
Agent Foundry is a local-first reliability and provenance debugger for AI agents. It shows why an agent behaved differently, where a run first diverged, what evidence and tools influenced its decision, and whether the behavior can be replayed.
Our Hack Day demo uses a simulated autonomous biology workflow based on the Moddik plate architecture. Deterministic sensor events become independently addressable evidence, are committed into verified hardware-state Merkle breakpoints, projected into Neo4j, and consumed by a local Liquid AI model. Agent Foundry then exposes the exact evidence route from sensor state to model decision and simulated intervention.
We also use PLAUD as an independent audio chain-of-custody channel for the team/demo session, preserving the original recording as an addressable evidence artifact.
The demo shows a controlled replay in which changing one sensor observation produces a different downstream decision, allowing Agent Foundry to identify the first divergent event.
Core idea: instead of asking only what did the agent answer?, Agent Foundry answers why did it answer differently, what evidence did it use, and can we reproduce the run?
Note: the demo URL is substituted with the GitHub repository link.
