Trust, rather than speed, is the main constraint when AI handles financial decisions.
2
A financial harness should combine flexible model behavior with deterministic code for calculations and tests.
3
The five trust principles are shared domain, consistency, evidence, accountability, and verification.
Summary
Francesco Galletta and Felippe Felisola Caso build a financial-analyst harness around the practical problem of budgeting a wedding. They argue that models are already fast enough to produce answers, while money-related work needs systems that make those answers understandable and dependable. Their design uses five principles. The user and model first establish a shared domain through an interview and an explicit plan. Calculations then move into code, where they can be run consistently and tested. Evidence records where inputs and decisions came from, including changes to code and assumptions. Accountability adds review gates at high-signal decision points, while verification brings in a separate perspective, golden examples, or outside reference points. The speakers also explain when a dedicated harness is worth building instead of adding more instructions to a general coding assistant. Their answer depends on the cost of mistakes, the number of users, and the need for reproducible results.
Galletta and Caso say that producing an analysis is no longer mainly a speed problem. People already have agents running background tasks, but money creates a high-stakes setting where users need confidence in the result. They distinguish reusable components such as traces, loops, graphs, observability, unit tests, and evaluations from the design principles that give those components a purpose. The workshop focuses on how a system should behave when people delegate financial work to agents. The goal is to make the boring work delegable without turning the result into a black box.
The harness starts by asking the user to describe a financial problem and then keeps asking questions until the plan has enough detail. In the wedding example, the user gives payment amounts, fixed costs, an unknown number of guests, a price per dinner, savings, and possible help from family. The harness is deliberately strict about vague inputs. It separates facts supplied by the user from assumptions made by the agent and from decisions that remain open. The resulting plan becomes a decision tree, with individual nodes that can be inspected and corrected without restarting one long conversation.
Long linear chats lose context and make correction difficult
The speakers describe ordinary agent chats as linear discussions that keep everything in one large context. After many turns, users often feel that earlier decisions have disappeared and want to start again. Their harness breaks the work into individual plan elements and supports branches for side questions or alternative reasoning. A user can correct one assumption, inspect its consequences, and return to the main plan. This structure also prevents a later sentence from contaminating the whole analysis. The harness guides the discussion toward a plan instead of immediately agreeing to an unclear request.
Caso and Galletta argue that arithmetic should not depend on the model generating a fresh formula in each session. A model can produce a formula that looks convincing while returning a wrong result, or it can give slightly different answers when asked the same question again. The harness identifies calculations and moves them into code that can be run repeatedly, versioned, and tested. They recommend ordinary software practices such as modular code, unit tests, end-to-end tests, and golden examples. The model can decide when to use the calculation tool, but the harness should prevent it from solving deterministic problems by parsing tokens in its head.
Evidence must show where inputs and decisions came from
Evidence means showing the receipts behind an analysis. A bank statement can be treated as a fact, while an agent's remembered information is an assumption and an unstructured thought is an interpretation. The system should preserve those distinctions and let a user trace a result back through its inputs, decisions, code, and data sources. The speakers warn that this also applies to deterministic code. A shared master file may change its assumptions without the analyst noticing, so code versions and process changes need traceability. A visible trail lets people investigate a number without reading every log or redoing every calculation.
The person using an agent remains responsible for its output, even when many agents are involved. The speakers reject both extremes: approving every tiny change and reviewing nothing. Their harness groups work into tasks and places review gates around important judgment calls, ambiguity, and consequential decisions. Tests can provide the gate for code, while a person can review a plan or choose between scenarios. The design should also support course correction. A user can change an input such as the expected number of wedding guests and see the plan reconsider its downstream consequences. Accountability is therefore built into the workflow rather than added as a final check.
Verification adds an independent point of reference
Verification gives the analysis another perspective without transferring responsibility to the reviewer. The speakers describe branches that review assumptions, challenge whether a plan seems reasonable, and then rejoin the main workflow with a consequence when the challenge finds a problem. Golden examples provide a reference for what correct data should look like, while outside sources can catch errors that are invisible inside the original dataset. They compare this with an experienced finance person noticing that a total feels wrong without checking every cell. The point is to create an intentional feedback loop that can expose bias, selection effects, and implausible results.
A dedicated harness is justified by risk and repetition
There is no fixed rule for how many harnesses a team should build. The speakers suggest considering how many people will use the workflow, how costly mistakes are, and how often a general-purpose harness produces the wrong result. A skill or prompt can be enough for a low-risk task where a person can inspect the output. A dedicated harness becomes more useful when reproducibility matters, when many users rely on it, or when a small amount of variability can affect an important decision. They also warn against spending more time sharpening the tool than using it. Repeated copying between chats and repeated manual steps are signs that a workflow may deserve its own structure.