A man studies a monitor showing friendly AI robots assisting with files, gaming, and online shopping.
Three weeks into any AI-agent experiment, you tend to find that the dull jobs go well and the exciting ones make you nervous. How-To Geek contributor Arol Wright's first-person piece, "3 genuinely insane things I've done with the help of AI agents," is mostly about that split. He describes handing an agent his messy folders, a game idea and a shopping cart. The "insane" part is mostly marketing. The useful part is how each task was fenced in.

This is one writer's anecdotal account. It doesn't name the file or shopping agent, and it has no logs or measured time savings. The tasks are real and common, though, and Microsoft's own developer documentation helps show what a sensible permission setup looks like.

A man studies a monitor showing friendly AI robots assisting with files, gaming, and online shopping. Experiment one: letting an agent sort Downloads and Documents​

Wright says Downloads is where files "go to be forgotten": years of installers, screenshots, receipts, drafts and duplicate PDFs. His agent didn't rename files at random. It first inventoried the folders, looking at file types, names and modification dates. In many cases it also opened documents to see what they were.

From that it proposed a folder structure, such as invoices, study material, installers, images and work drafts. His example is a file called scan0043.pdf, which stops being a mystery once the agent reads the first page and sees a utility bill.

The workflow matters more than the result:

  1. Ask for a plan before anything moves.
  2. Review and correct the plan.
  3. Only then let the agent execute.
  4. Have it move files rather than delete them.
  5. Have it flag duplicates and old installers for you to review instead of erasing them.

Wright says this turned an afternoon of dragging and dropping, or possibly several days, into a plan review and a few corrections. He also says the agent applied naming conventions more consistently than he does by hand. Treat those as his estimates, not benchmarks.

Where the caveats sit​

"Move, don't delete" removes one kind of irreversible error. It doesn't stop a file being filed in the wrong place, which can be nearly as bad when you need that tax document next April. Sensitive or ambiguous files deserve your own eyes. A backup before you start is cheap insurance.

Access scope is also worth checking. OpenAI's help documentation for ChatGPT Work says the desktop app can use local files only when you open a folder or project and grant access. It advises granting access only to the files the task needs. It also says messages and task context may be stored in the cloud even when the work runs locally. Web and mobile versions can't reach files on your computer directly. Wright doesn't say which agent he used, so this is context, not a description of his setup.

If you try this, point the agent at a copy of the folder, or at a single subfolder, rather than your whole Documents tree.

Experiment two: vibe-coding meme games​

Wright says he has made a couple of spare-time games with Codex, including a horror game. He defines vibe coding as describing what you want in plain language and iterating by conversation instead of syntax. A meme game starts with one joke and one mechanic, such as dodging falling objects or clicking as fast as possible. The agent then produces a single page of HTML, CSS and JavaScript that runs in any browser, with no frameworks or build tools. Testing means opening the file.

The fun is in the revision loop. The first version is rarely right: collisions feel off, difficulty ramps too fast, or scoring makes no sense. You say the game is too hard in the first ten seconds, or ask for screen shake when the player loses, and the agent revises it. If the agent can run the page and check for errors itself, it can catch a broken script before you see it.

Wright is candid about the limits:

  • Vibe-coded projects get messy as they grow.
  • You learn little programming from code you never read.
  • Agents lean on simple shapes and emoji unless you supply art.

He says an idea that once needed a weekend now takes an evening. He also admits he stops playing his creations after a couple of days. That is a reasonable verdict. A playable prototype isn't proof of correctness, security, accessibility or maintainability.

What the permission controls look like in VS Code​

Coding agents don't only write text. They edit files and run commands. Microsoft's Visual Studio Code documentation describes the problem and its fixes. It says agents in Visual Studio Code can edit files, run terminal commands, and call external tools. It adds that VS Code uses two layers. Approvals determine whether an action runs automatically or requires your confirmation. Sandboxing restricts the file system and network resources that agent terminal commands can access, even after the commands are approved.

The docs list the main settings as chat.permissions.default for permission levels, chat.tools.eligibleForAutoApproval for tools, and chat.tools.terminal.autoApprove for terminal commands. The fetched documentation describes three permission levels for Copilot sessions:

LevelBehavior
Manual permissions (default)Uses your tool, URL and terminal approval settings; anything not auto-approved asks you first
Assisted permissionsAn LLM judge assesses each tool call, and calls it doesn't approve ask you
Allow allRuns all tool calls without confirmation

The documentation warns that Allow all and Autopilot skip confirmation for potentially destructive actions, including file edits, terminal commands and external tool calls. It says to use them only when you trust the workspace and understand the risks. It also says common read-only commands run automatically by default, while risky ones such as rm and del require approval.

Microsoft's own documentation is blunt about the limits of rule-based approval. It calls terminal auto-approval a best-effort convenience, not a security boundary, and says obfuscated commands can evade matching. One of the results I found says that for prompt-injection worries, sandboxing or a dev container is the stronger choice than auto-approval rules alone.

Two cautions apply to sandboxing:

  • Sandboxing is not a virtual machine or user-account boundary, and a sandboxed process still runs on the execution host under your account.
  • Platform availability varies in the documentation I saw. The page fetched for this story labels sandboxing as Preview on macOS, Linux and WSL2 and Experimental on Windows. Another Microsoft page describes support on Windows, macOS and Linux, and a third-party mirror lists only macOS and Linux, including WSL2. Check the current docs for your platform and VS Code version before relying on it.

None of this describes Codex's own settings or Wright's exact workflow. The VS Code documentation does show that Codex can be used through VS Code, with its own presets. Default Permissions asks before using the internet or touching resources outside the workspace, while Full Access removes those limits.

For a throwaway game, a practical routine is:

  1. Keep the project in its own folder.
  2. Ask for small changes.
  3. Run the page after each change.
  4. Read the code before reusing or expanding it.
  5. Keep the agent on manual approvals unless the project is genuinely disposable.

Experiment three: letting an agent shop​

The third experiment carries the most obvious risk. Wright says browsing agents can now move through online storefronts like a person: searching, comparing listings, reading reviews, picking options and going through checkout. The appeal is tedium reduction. For something like a compatible charger or a replacement part, it can replace a dozen open tabs with a shortlist of tradeoffs in price, shipping time and seller reputation.

His boundary is that the agent prepares and you pay. He says well-designed agents stop before the final step, fill the cart, show the item, total and payment method, and wait for clear confirmation. He also says card numbers and passwords should be entered by you or handed over through a password manager, not typed by the agent in plain text. His advice is to start with low-cost items from reputable stores and check every order summary.

Note that "well-designed agents" is a generalization. His article names no agent, store, product or completed order, so it can't tell you that every shopping agent pauses before paying. Check the behavior of the specific tool you use.

Why the review step isn't optional​

Wright lists misread listings, wrong variants and misleading page text as risks. Independent security guidance backs the last one up. OWASP's GenAI guidance describes indirect prompt injection, where an LLM accepts input from external sources an attacker can control, such as websites or files. One of its listed examples is a rogue instruction embedded on a visited website exploiting an e-commerce plugin, leading to unauthorized purchases. That description comes from OWASP's older LLM Top 10 entry, but the mechanism still applies to browsing agents.

OWASP's mitigations map closely onto Wright's advice:

  • Give the model least privilege.
  • Require human approval for privileged actions.
  • Treat the model as an untrusted user and keep final decisions with the person.

OWASP also notes there is no fool-proof prevention, because LLMs don't separate instructions from external data. A confirmation screen is a safeguard, not a guarantee. A compromised model could still present misleading information, so read the cart contents, quantity, shipping address and total yourself.

OpenAI's computer-use documentation, which concerns its API tool and not Wright's agent, makes a related point. Asking for confirmation depends on the agent actually calling the confirmation tool. Applications that need a guaranteed approval step should restrict the browser to resources that can't make purchases, or use a controlled runtime.

What the three stories have in common​

Each task rewards the same pattern, whatever tool you use:

  • Narrow access. Give the agent one folder, one project directory or one low-stakes store, not your whole machine or wallet.
  • Plan first, act second. Ask for a proposed plan, folder map, diff or cart before anything changes.
  • Recoverable actions. Prefer moves over deletes, work on copies, and keep backups.
  • A human on irreversible steps. Deleting files and spending money should need your explicit yes.
  • Skepticism toward untrusted content. Web pages and documents can contain instructions aimed at the agent, not at you.

Wright's closing line is that agents should be treated as fast assistants, not autonomous decision-makers. That is a sensible rule, and for Windows users it comes down to settings. Check what your agent is allowed to touch, which approvals it asks for, and whether you've quietly switched on an "allow everything" mode because the prompts got annoying.

Approval fatigue is a real concern. Microsoft's own trust-and-safety documentation notes that repeatedly approving commands can make you pay less attention to what you're approving. The usual fix is to narrow what the agent can do in the first place, not to click through faster.

The adventures in this piece are modest: tidy folders, a silly game and a shopping shortlist. Wright's own account is that agents work best on tedious, low-stakes tasks where mistakes are recoverable, and he doesn't claim more than that.

 

References

  1. 3 genuinely insane things I've done with the help of AI agents How-To Geek 2026-10-22T21:00:17+00:00
  2. ChatGPT Work and Codex | OpenAI Help Center help.openai.com
  3. Manage approvals and permissions code.visualstudio.com