Research project / live system
Build Guardian
A local-first lab that replays agent workflows under controlled perturbations and measures where their state begins to drift.
why it exists
A single successful run says little about whether an agent workflow is stable.
how it works
The harness holds workflow and seed steady, changes one condition, then compares role-aligned state and outcomes.
how it looks
A control room connects the workflow graph, pass inspector, state diff, replay oracle, and fragility score.
ModesFull, checkpoint isolated, anchored, drift
OracleEight pure replay perturbations
MetricRole-aligned drift, abstention, latency, fragility