LongHorizon-Harness keeps agent state only after an audit

LongHorizon-Harness splits a long task across a manager, executor, and auditor. The executor starts each round with fresh context, and only results checked independently in the environment enter persistent task state. Why it matters: Long tasks break when an agent's memory records guesses as progress. This design makes verification a condition for carrying work into the next round.

Try this: Run one low-risk task with and without an audit gate. Compare the final files, the state that survives a reset, and the checks recorded for each run.