Chronicle addresses the problem that LLM agent failures are hard to reproduce because model calls are non-deterministic and tools read changing state. It records agent runs at non-deterministic boundaries (model calls, tool calls, routing decisions) as immutable envelopes. Its core operation, cut-point replay, lets a developer choose which boundaries to serve from the recording and which to execute live with new code, turning a recorded incident into a regression test that can run in continuous integration.
On a benchmark of six recorded failures, recording adds only 23 microseconds per boundary crossing (0.008% of a typical 300 ms model call). Full replay reproduces the exact run with zero model calls and is bit-stable across 20 repetitions. Cut-point tests correctly fail on faulty code changes and pass on guarded fixes and benign edits. In a mutation study, cut-point tests caught every mutant that allowed the recorded unsafe action, while a baseline that stubs every boundary caught none.
This matters because it makes flaky LLM agent failures reproducible and testable, enabling regression testing without costly model calls and providing better fault detection than full stubbing approaches.