A financial analyst team wants their Foundry-hosted agent to let users upload a spreadsheet of quarterly transactions and ask questions such as 'what was the average transaction size in March,' with the agent performing the actual calculation on the uploaded data rather than guessing from memory. Which Foundry Agent Service capability should the team enable?
Select an answer to reveal the explanation.
Short Explanation
What the analysts actually want is real math done on real numbers, not a plausible-sounding guess. Giving the agent a code-execution tool lets it treat the uploaded spreadsheet like an actual dataset — read the March rows, run the average, hand back the true figure. Just telling it in the instructions to be careful with numbers doesn't change what's happening under the hood: it's still predicting the next likely words, not performing arithmetic, so precision isn't guaranteed. Training it ahead of time on sample spreadsheets teaches it what transaction data tends to look like, but that's pattern recognition, not a way to crunch this specific file's numbers months later. And pulling in relevant snippets of text is great for finding a passage to quote, but it was never built to add up a column of figures accurately. When the ask is compute this exactly, you need something that can actually run the computation.
Full Explanation
The correct answer is A. The code interpreter tool lets a Foundry agent write and execute code against an uploaded file, so when a user asks for the average transaction size in March, the agent can actually run a calculation over the real numbers in the spreadsheet and return an exact result rather than an estimate produced purely from language generation. Option B is incorrect because a system message can instruct the agent to be careful, but the underlying model still generates its answer through language prediction rather than genuine computation, so exact figures like an average across many rows are not reliably guaranteed. Option C is incorrect because fine-tuning trains the model on example patterns it has seen before, which helps with style or typical structure but does not let the model compute a precise statistic on a new spreadsheet's specific values months after training. Option D is incorrect because retrieval-augmented generation is built to pull relevant passages of text for the model to reference, not to execute numeric operations, so it could surface rows from the file but would not reliably perform an accurate calculation across them.