ErosionBench is a benchmark that tests whether multi-agent systems hold to their original instructions over many sessions, not just the first one. The core finding is that alignment, meaning how closely agents follow their assigned rules and values, decays over time as agents hand work to each other and accumulate context. For teams designing enterprise agent products, this is a named, measurable problem, not a theoretical one.
The benchmark includes working scenarios that simulate real enterprise conditions, so teams can run their own systems against it. It also ships a reference defense, giving product teams a starting point for reducing drift rather than just measuring it. The practical value is that "does this agent still behave the way we set it up?" now has a test with a standard.
For designers and product leads, the interaction this exposes is the gap between the brief a human sets at the start and what agents actually do several sessions later. A human delegates work, walks away, and trusts the original constraints still hold. ErosionBench shows that trust erodes in ways that are hard to see without deliberate measurement, and that the erosion compounds when agents pass work to each other.