Francesco Galletta and Felippe Felisola Caso use a wedding budget to demonstrate a financial-agent harness at AI Engineer. They work at Monzo but explicitly say this is not a Monzo product demonstration; the workshop prototype is intentionally rough.

Trust is a design problem

  • Shared domain: the agent interviews the user before solving anything. A plan breaks the decision into nodes and labels what the user stated, what the agent inferred and what remains open.
  • Consistency: calculations that must repeat should live in versioned, testable code rather than be regenerated as model prose. A changed formula should invalidate confidence in downstream numbers.
  • Evidence: every input and decision needs a path back to its source. A guess, a user statement and a bank record have different evidentiary status; even a trusted “master file” can change its rules.
  • Accountability: humans retain consequential choices. Review a few high-signal gates rather than approving every intermediate step or surrendering the whole workflow to an agent.
  • Verification: compare against a second reference point, not only the data and examples that produced the answer. A challenge must feed back into the plan or be explicitly dismissed.

What the wedding demo actually shows

  • The initial prompt includes three payments, a known first payment, an unknown middle payment, fixed costs and a per-guest dinner cost. The speakers deliberately add vague or changing inputs so the agent must ask questions before calculating.
  • A seeded plan graph lets participants inspect provenance, change a guest-count assumption, run tested calculation nodes and branch a review discussion before merging a correction back into the main plan.
  • The workshop code illustrates a way to implement these principles. The presenters do not claim it is a production-ready financial tool, prove it works for every user’s finances, or show a tested Monzo deployment.
  • Their rule for building a dedicated harness is practical: invest when a repeated, consequential workflow needs reproducibility and a generic prompt/skill keeps creating avoidable errors. Don’t build a new framework for every minor task.

“Agents face no real-world consequences. You do.” — a saying Galletta cites without attributing its original author