Give the agent a concrete deliverable
Specify the input, the output format, and what counts as success. For example, ask for a cleaned CSV and a reusable script, or an interactive artifact with a short explanation of how to use it.
Run, inspect, and revise
Agent can read and write workspace files, execute code, and browse for information. It can use the results of one tool call to decide the next step. Tool availability depends on your plan and the task.
Check the work before using it
Ask the agent to run relevant checks and report what it actually verified. Review generated code and calculations. Passing a test is evidence for that check, not a guarantee that every possible input is handled.
Understand an Agent request
Each submitted Agent request, including a follow-up, counts toward the plan's Agent allowance. The internal tool calls within it are not separate Agent requests. Agent work also consumes rolling chat usage, which varies with the model, history, and task.
Keep the output you need
Download the files you want to retain. Cloud Vault is optional; consult the storage guide for how local history and encrypted cloud storage differ. A sandbox is an execution workspace, not a backup service.
A brief to try
Inspect this sales CSV for missing dates and invalid amounts. Write and run a script that summarizes revenue by month. Return the script, a cleaned CSV, and a report explaining any rows you excluded.
Review the resulting files and the reported checks. The brief illustrates a workflow; no completion time, cost saving, or competitor parity is implied.