I am Jarvis Nelson Osorio, also known as Jarvis Osorio, a technology entrepreneur and innovator based in Chile. I work on AI systems, DAO design, writing and art. As soon as an agent can touch money—rebalance a book, prepare a transfer, reconcile an invoice, or construct a transaction—the most dangerous phrase in the architecture is "it checked itself."
I originally shared this argument on LinkedIn in Why Finance AI Agents Need Independent Validation: Inside Screwtape.ai. The product landscape will move. The design rule should not: the process that proposes a financial action must not be the only process that blesses it.
The self-grading problem, with a price tag
In many multi-agent demos, one model writes the plan and the same family of model "reviews" it. That feels like governance. In numerical and contractual work it is often an echo. If the first pass misreads a margin, a date, or a clause, the second pass is statistically likely to find the error reasonable. In a chatbot, that is an awkward paragraph. In a ledger, it is capital.
Finance does not forgive a confident hallucination the way a brainstorming session can. A single unverified write can be very expensive. So the standard I argue for is simple and strict: before an agent executes a trade, issues an invoice, or updates a balance, an independent layer should return a clear verdict—pass, fail, or retry—against constraints you actually wrote down.
Consensus can be a false friend
Teams sometimes add more agents and take comfort when they agree. I have also written about the opposite case: several agents can share a flawed assumption and call it confidence. In a financial swarm, agreement should be scored, and "we all said the same number" should raise a question: did they share a bad input? An independent check that flags false-consensus risk is more useful than another teammate trained the same way.
Receipts, not vibes
If a regulator, a partner, or your future self asks why an autonomous system moved money, "the model thought it was fine" is not a record. A signed, inspectable receipt of the checks performed before execution is. This is institutional design, the same family of ideas I develop in Breaking the Centralized Order: visible rules, separated powers, an audit trail that does not depend on anyone's memory.
What "independent" has to mean
Independence is not a second tab with the same prompt. It is a different evaluator, a deterministic test, or an external service that cannot be talked out of a rule. It sits beside the executing agent, not inside its self-talk. It should be allowed to say no.
You can build that layer yourself or use a dedicated judge service; the original LinkedIn piece looks at one such pattern. Either way, do not let a finance agent hold the pen and the gavel. Let it draft. Let something else, with a record, decide whether the draft may become a transaction.
Where to go next
This expands on ideas I originally shared on LinkedIn in Why Finance AI Agents Need Independent Validation: Inside Screwtape.ai.
More essays, books and profiles: jarvisosorio.com and the hub Jarvis Nelson Osorio — articles, books, videos and profiles. The same separation of powers is the subject of Breaking the Centralized Order.
More from Jarvis Osorio: https://jarvisosorio.com
Top comments (0)