What bank reference data taught me about letting agents write

Before agents, I worked on Party Master, an entitlement-sync service on a wealth platform serving 10M+ users, where clients manage billions and every downstream system trusts yours to be right. Reference data is the quiet tier of a bank: nobody notices it until it's wrong, and then everybody does. It turned out to be perfect training for the agent era.

The disciplines transfer one-to-one. Reconciliation: never trust a write happened because you sent it, verify against the source, continuously; my reconciliation scripts auto-deactivated users that upstream sources no longer justified, because stale entitlements are silent risk. Audit: every mutation attributable, timestamped, and immutable, because 'what changed and why' is the first question in every incident. Idempotency: replays and retries must converge, not compound.

Now replace 'upstream feed' with 'agent.' An agent writing into a system of record is a very fast, occasionally confused upstream source. The same machinery applies unchanged: verify-after-write instead of trust, append-only audit of every agent action, idempotency keys so a retried step can't double-execute, and reconciliation sweeps that catch drift between what the agent believes and what the system holds.

The industry is currently re-deriving these lessons and calling them agent safety. Banks derived them decades ago under the name operational risk. If you're designing agent write-paths, hire the discipline, not just the model, the boring parts of fintech are the load-bearing parts of agentic systems.