Francesco Galletta and Felippe Felisola Caso told AI Engineer to stop chatting with spreadsheets and build a financial analyst harness you can actually trust, and I think they nailed the bottleneck. Speed is not the problem anymore. Trust is.
Their workshop used a deliberately small example, budgeting a wedding with fixed DJ costs, variable catering and three payments, to show why a blank chat box fails on money. Galletta said the job of an analyst spans research for banks to personal finance, but the failure mode is the same: the model does not share your definition of a term, so you must align on language first and only then let tools do arithmetic. I buy that sequence, and the concession is fair too. If you want to throw 100 agents at a problem and hope, you can, but you will be grading vibes, not verifiable math.
What had to be true before this talk could exist is a few years of deliberate internal adoption at Monzo. Felisola Caso, who moved to the UK from Brazil and worked at Nubank before Monzo, said the team refuses token maxing without concrete business value and builds AI to make the data team itself better. Galletta, now based in Madrid after seven years in the UK, framed the same history from the personal side, automating what he cares about rather than wiring an OpenAI client to everything. That shared context explains why they demoed harness principles over product tours.
What they demoed was not a product but a harness with five stacked principles: shared domain, consistency, evidence, accountability and verification. The affected system is not a customer-facing app; it is the internal analyst workflow where a number becomes a decision. The attacker requirement here is local, not remote. No exploit is needed for harm, just a confident wrong total copied into a payment plan, which is why Felisola Caso stressed traces, loops, unit tests and observability are components, not trust itself.
What still has to happen is the hard part they left explicit. Consistency must come from tool calling and code, not language model arithmetic. Evidence must show receipts you can follow back. Accountability must keep a human responsible for a swarm, without demanding line-by-line diff review that no one does. Verification needs a separate perspective to catch bias. The workshop harness is intentionally overengineered and glitchy, they warned the UI is not slick and something will break, and that honesty helps. Outside this room, measurement is tightening too: the Vals AI Finance Agent Benchmark v2 expanded its harness to add calculator and price_history tools on top of v1.1 and now reports mean-of-runs across three runs with standard error, a reminder that single-shot demos are not the metric anymore.
