Some tasks — reconciling a schedule of locations, checking premium math on a proposal, reshaping an exported spreadsheet — are more reliably done in code than in natural language. Cara can execute Python inside a sandboxed environment attached to your workspace and use the result to continue the task.
Code execution is always visible. Every run appears in the workspace’s Execution tab with the exact code, inputs, stdout, stderr, and any files produced.
When Cara uses it
Cara reaches for code execution whenever a task benefits from deterministic computation, including:
- Tabular reshaping — pivoting a schedule of locations, joining two spreadsheets, deduping rows, normalizing carrier names
- Verification — recomputing a premium sum, checking that TIV rolls up correctly, confirming that ACORD 125 totals match the underlying schedule
- File conversions —
.xlsx → .csv, extracting specific sheets or columns, filtering by date
- Small parsers — pulling structured data out of an unusual PDF layout or a text dump
You can also ask directly: “Use code execution to…” or “Run Python to check that…”. Cara will draft the code, execute it, and quote the result back to you.
What it can access
The sandbox has read access to any file you’ve attached to the workspace and any artifact Cara has generated in that workspace. It does not have network access, and it does not have direct access to your integrations or knowledge base — Cara pulls that data via her regular tools first, then hands it to the sandbox.
Cara treats code-execution output as evidence, not as truth: she still narrates her reasoning in the Messages tab and cites the specific run when relevant.
Reviewing a run
Open the Execution tab and click any code-execution step to expand it. You’ll see:
- The Python source that ran
- The input files referenced
stdout and stderr- Any new files produced (they’ll also show up in the Artifacts tab)
If a run failed, Cara automatically retries with a corrected version and keeps both attempts in the history. 
