Lapse

Lapse is a dynamic race-condition detector for AI agents that finds bugs in the side effects an agent produces through its tool calls. It catches the silent corruption that happens when an agent issues concurrent or retried tool calls against the same external resource, including double charges, lost updates, duplicate records, and other order-dependent failures that pass every sequential test and slip past LLM-based evals.

Beyond statically flagging conflicts in a recorded run, it actively provokes latent races by forcing alternate interleavings and then reproduces any failure deterministically with a minimal, seed-replayable schedule. Its verdicts are purely algorithmic rather than a model's opinion, and it plugs into existing workflows through a pytest plugin, a GitHub Action, and capture adapters for popular agent frameworks.