Agent Inputs & Files Copied!

Every agent runs on inputs: the files it processes and the values that change from period to period. This article covers the input types, how to prepare files that build and run cleanly, and how inputs behave differently at build time versus run time.

Input Types Copied!

An agent's inputs are defined while you build, and each one has a type:

File A spreadsheet or other document the agent processes, like a trial balance export or an invoice register. The input type you'll use most.

Date A date you set each time the agent runs, like the period-end date. Use it anywhere your instructions need a date that changes every period, so you never have to edit the agent itself.

Text A text value you set each run, like an entity name or a threshold amount. Same idea as a Date input: the value changes per run, the logic doesn't.

Date and Text inputs are what keep an agent reusable. If you find yourself wanting to hardcode "March 2026" or "$5,000" into a step's instructions, that's usually a sign it should be an input instead.

Referencing Inputs in Your Instructions Copied!

When a step needs to use an input, refer to it by its exact name, wrapped in quotes. If your agent has a Date input named 'Period End' and a File input named 'Invoice Register', a step's instructions might read:

    Filter 'Invoice Register' to rows where the 'Received' date is on or
    before 'Period End'.
  

The quoted name tells the AI exactly which input you mean, the same way quoting tab and column names does. Naming inputs clearly pays off here: 'Invoice Register' reads better in instructions than 'File 1', and it tells whoever runs the agent next period exactly what to upload.

Build Files vs. Run Files Copied!

The file you build with is a design tool, not the real thing. At build time, the AI reads your file to understand its structure and write the script. At run time, the script processes the file you provide for that period. Once the script exists, the build file's job is done.

Build with a trimmed sample rather than a full export. Large files slow the AI down during building and can cause timeouts, and Transform samples very large files automatically, which can strip formatting. A few hundred representative rows is plenty.

"Representative" is the key word. Your sample should include the edge cases the real file will have: the odd date formats, the blank cells, the vendor names with special characters. A script built against clean, uniform data will stumble the first time the real file isn't.

Run files can be much larger. The sample is only for designing the script; once the script exists, it can process the full export each period.

Preparing Your Files Copied!

A few minutes of file cleanup prevents most build and run problems. Before uploading a file:

Use .xlsx or .csv for data files. PDFs are supported by the Extract skill for pulling data out of unstructured documents, but data processing steps expect tabular files.
Put headers in the first row of each tab, and keep one table per tab. If a sheet has stacked tables or title rows above the headers, split and clean it first.
Remove images, comments, and notes. Screenshots and cell comments can break the file preview.
Break external workbook links. Links to other workbooks aren't supported. In Excel: Data → Workbook Links → Break All.
Resave the file if it misbehaves. Some system-generated .xlsx exports need to be opened and resaved in Excel before Transform can process them. If a fresh export causes odd errors, try that first.

Changing Inputs After You Build Copied!

Inputs aren't locked once the agent is built. You can replace the file attached to a File input, update Date and Text values, and rename inputs as your process evolves. When you swap a build file, Transform rechecks the steps that use it, so you'll see right away if the new file's structure breaks anything.

If an input is no longer referenced by any step, you can delete it. Unused inputs are worth cleaning up, since anyone running the agent has to look at every input it asks for.

What providing inputs looks like at run time, including replacing files before a run starts, is covered in Running an Agent.