Curriculum previewThis is not your assigned passport. No identity, answers, or completion progress are stored here.How to start your passport
Mission / AI Agent User
02 / 0240 active minutesdata

Use A Safe Agent Loop After The Manual Baseline

You can bound an agent task, control context and commands, review the diff, verify the result personally, and reject a plausible but inferior suggestion.

Work herelocal / powershell or zsh or bash
  1. 1Understand
  2. 2Do
  3. 3Check
  4. 4Submit
Learn firstConcept, example, trap and recovery

Outcome#

You can bound an agent task, control context and commands, review the diff, verify the result personally, and reject a plausible but inferior suggestion.

Concept#

An agent can read files, edit code, execute commands, and publish changes. It is not merely a model or chatbot, and fluent output does not transfer responsibility away from the account owner.

Worked Example#

The fixture is corrected, the canary remains unchanged, the bounded verifier passes, and the submission records human review rather than a transcript.

A correct example uses these decisions:

  • What is a safe agent task? One bounded outcome with explicit files, constraints, and verification.
  • Who verifies the final diff and tests? You do, even if the agent reports success.

Common Trap#

Granting a broad task, sharing protected data, or accepting a claimed test result without checking the diff and rerunning it.

If Blocked#

Start a new bounded agent thread if context has become inconsistent. Return to the clean baseline when edits spread outside the fixture. Use the manual versus agent lab for recovery.

Useful references:

Understand Before Accepting AI Output#

You must understand the changed behavior, tests, files, commands, data sent, service used, and likely cost. Passing output without this explanation is not a pass.

Do the mission

One bounded attempt

Pass: 80% + every safety item
Action

Give an agent one bounded fictional task, inspect every changed file and command, reject one unsuitable suggestion, then run the checks yourself.

Expected

The fixture is corrected, the canary remains unchanged, the bounded verifier passes, and the submission records human review rather than a transcript.

Separate practice space

Open this mission's safe practice files

The launcher creates or validates a separate exercise folder. It stops rather than replacing an existing unrelated folder.

Run only the commands for your machine

Step 1windows / powershell
git status --short -- workspace/agent_task

Expected: Only workspace/agent_task/storage-plan.md is modified.

Step 2macos / zsh
git status --short -- workspace/agent_task

Expected: Only workspace/agent_task/storage-plan.md is modified.

Step 3linux / bash
git status --short -- workspace/agent_task

Expected: Only workspace/agent_task/storage-plan.md is modified.

Step 4windows / powershell
git diff -- workspace/agent_task

Expected: Only the intended storage-plan change is shown; canary files are untouched.

Step 5macos / zsh
git diff -- workspace/agent_task

Expected: Only the intended storage-plan change is shown; canary files are untouched.

Step 6linux / bash
git diff -- workspace/agent_task

Expected: Only the intended storage-plan change is shown; canary files are untouched.

01What is a safe agent task?Safety-critical
02Who verifies the final diff and tests?Safety-critical
Before submitting

The launcher will copy only this mission's declared synthetic practice files into the public submission. Review those files as well as these answers.

Open my real passport first
Preview onlyThis page can be read, but it cannot store or submit an attempt.