CrewAI adds spreadsheet queries to agent inputs CrewAI adds spreadsheet queries to agent inputs

The preview is a map, not the dataset

The change follows version 1.15.37, released on November 7, which introduced spreadsheet reading in crewai-files. The newer release connects that machinery to URLs in an agent’s kickoff message or task description. With the optional spreadsheets package installed, the input supplies the file location and the prompt gains a summary of columns, data types and sample values. Two tools then separate orientation from calculation. inspect_spreadsheet describes the available sheets and columns. query_spreadsheet can retrieve pages of rows, count values, group records, aggregate figures or calculate correlations between numeric columns. It can apply up to five filters before those operations.

A tool call reaches beyond sampled rows

The earlier spreadsheet implementation separates the material shown to the model from the underlying table. A compact preview helps the agent choose an operation without requiring every cell to fit into its prompt. The query layer then works against the loaded data. That distinction matters for a sales total or a count of support tickets: a handful of example rows describes the structure, while an aggregate needs the relevant records. The model still chooses the column, operation and conditions. The feature changes how evidence is retrieved and computed, rather than making a preview itself sufficient evidence for an answer.

The input determines which links are opened

CrewAI’s new input-handling code records caller-provided task text before planning or conversation history is added. Spreadsheet links produced later by the model or another tool are not automatically treated as caller input. There is an important boundary here: if an application explicitly passes a previous model response into a new agent’s input, that text counts as input. The documented downloader refuses private and internal network destinations. It also uses a shared deadline across spreadsheet downloads. This is a separate budget from the agent’s own execution-time setting, so it belongs in the application’s handling of slow or unavailable files.

The release includes checks for misleading file results

The accompanying tests cover a web page served at what looks like a CSV address, an expired download budget and more URLs than the run accepts. A page in place of a file is reported as unreadable instead of becoming spreadsheet data. Another regression test uses two different files with the same name. It verifies that the second answer comes from the second file’s rows, rather than a cached answer for the first one. These cases make the distinction between locating a file, reading its contents and returning a result explicit in the implementation.