How to Run a Safe First AI Agent Experiment

Use copied files and four observable stages—read, plan, act, review—to test an AI agent without external logins or irreversible actions.

Do not begin your first Agent experiment with an external login or production data. Prepare a recoverable copy and separate the run into read, plan, act, and review: understand the material, inspect the plan, authorize only bounded changes, and verify the before-and-after difference.

An Agent is useful because it can continue across several steps. That is also where its risk comes from. If the first experiment includes email, cloud storage, and write access, you will struggle to tell whether a failure came from the task, the tool, or the approval design.

Use one fictional text file. Preserve the original and make a copy named practice-copy.md for the Agent to edit.

Read: inspect the material without changing it

Read practice-copy.md only. Do not modify it.
Restate the document’s purpose and current structure.
Identify three possible improvements and anything you cannot confirm.

Check that it read the correct file and did not present guesses as facts. If it misunderstood the material, do not proceed.

Plan: inspect actions before execution

Propose a revision plan. Name each section you would change and explain why.
Constraints: do not add external information, modify the original file,
or create additional files.
Wait for my confirmation before making changes.

OpenAI’s Work guidance emphasizes that users can review progress, change direction, and approve important actions. Guidance for agentic models also recommends defining autonomy and approval boundaries. OpenAI: ChatGPT Work and Codex OpenAI: Model guidance

The official materials support bounded autonomy and human supervision. The read, plan, act, review sequence is my first controlled experiment, not an official OpenAI workflow.

A plan is not a ceremonial status update. Remove steps that exceed scope, then approve only the smaller plan you understand.

Act: execute only the approved change

Execute only items 1 and 2 from the approved plan in practice-copy.md.
Do not use the network or modify any other file.
If required information is missing, stop and report it.

Watch whether the files, tools, and actions remain inside the boundary. If the Agent asks for more access, do not approve merely because the task is already in progress. Check whether the added permission is necessary.

Review: inspect differences, not only the completion message

Ask the Agent to list the changes it made, then compare the original and the copy:

“Completed” is a status statement, not acceptance evidence.

Save the experiment record

Controlled Agent experiment card

Original material and copy:
Was the Read summary correct?
Original Plan:
Steps I removed or constrained:
Act I approved:
Tools and files used:
Differences found during Review:
Can the change be reversed?
Actions that remain unauthorized next time:

You are finished when every stage has observable evidence and the original file remains intact.

The next lesson separates four controls that appeared in this experiment: permissions define what is possible, approvals decide whether this action may proceed, stop conditions interrupt abnormal behavior, and recovery handles mistakes.

References