Lethe
An open benchmark for measuring how agent behavior changes during extended operation. The question is not only whether an agent can finish a task, but whether it still follows the same objective later.
The missing question
Short benchmarks can show that an agent knows how to perform a task. They say less about what happens after repeated decisions, interruptions, and accumulated context. In practice, useful agents can gradually lose the original objective while still producing plausible-looking work.
What Lethe measures
Lethe runs repeatable scenarios and checks changes in objective fidelity, decision consistency, behavior, and final task outcome. Results are grounded in external checks wherever possible instead of trusting the agent to grade itself.
Why verification matters
A confident completion message is not proof that the requested state exists. The benchmark therefore separates what the agent says from what the environment can verify. This makes silent inversions and incomplete work visible in the report.
Current state
The public repository contains the harness, scenario format, examples, and current research notes. The work remains active, and claims about comparative model performance are published only when the underlying run evidence is ready to inspect.