Giant smartphone with a glowing checkmark towers over a icy landscape, connected by a glowing line to a laptop on a small stool.

Quick answer

To approve AI coding agent changes from your phone safely, only approve small, reversible changes you can understand on a small screen, and send everything else back with a question or leave it until you are at your desk. Tools such as Claude Duet, which now supports Claude Code sessions, let you watch progress, approve proposed changes and send new instructions from your phone. That is handy. But a quick tap is still an approval. Set limits before you walk away: work on a separate branch, block risky commands, keep secrets out of the agent’s reach, and make sure every change can be undone.

Key takeaways

  • Phone approval is convenient, but you see less on a small screen.
  • Approve small and reversible: text edits, styling, tests, new files on a branch.
  • Wait for your desk for deletions, payments, database changes, deployments and anything touching passwords.
  • Work on a branch so nothing reaches your live site without a proper review.
  • Ask before you approve: “What will this change, and how do I undo it?”
  • Keep API keys out of chat and out of the agent’s view.

📘 Main article: This guide supports Claude Silicon Leak: A New Model Family, OpenAI Cactus and Qoduck. It turns the new Claude Duet and Claude Code mobile support into a simple, safe approval routine.

AI coding agents now keep working while you make a cup of tea. In WorldofAI’s latest AI news video, the presenter explains that Claude Duet can now manage Claude Code sessions, so you can monitor a session, approve proposed changes and send extra instructions from your phone, whether the session runs on your computer or in the cloud.

This guide answers one question: how do you approve an AI coding agent’s changes from your phone without approving something you shouldn’t? For the rest of this week’s news, see our Claude Silicon leak round-up.

What does approving agent changes from your phone mean?

It means the agent pauses before an action and asks permission, and you answer from a phone app instead of your computer. Coding agents such as Claude Code usually ask before they edit files, run commands or install packages, depending on your settings.

From your phone you can usually do three things:

  • Monitor what the agent is doing and how far it has got.
  • Approve or reject a proposed change or command.
  • Send instructions, such as “stop and explain” or “use the existing button style”.

If you are new to agents that act on their own, our article on Claude Duet as a proactive AI agent explains the basics.

Which changes are safe to approve on a phone?

Safe changes are small, easy to read and easy to undo. If you cannot explain the change in one sentence, it is not a phone approval.

  • Copy and text changes on a web page.
  • Styling tweaks, such as colours, spacing or fonts.
  • Adding or updating tests.
  • Creating new files on a separate branch.
  • Running read-only commands, such as listing files or running tests.

⚠️ Warning: Do not approve from your phone anything that deletes files, changes a database, deploys to your live site, touches payment settings, edits passwords or API keys, or installs unfamiliar packages. Those can be hard to reverse and deserve a full screen and a clear head.

What should you set up before you leave your desk?

  1. Use version control. Make sure the project is in Git, so every change can be rolled back. Ask the agent to commit after each step.
  2. Work on a branch. Tell the agent to work on a new branch, never directly on your main or live branch.
  3. Set permission limits. In Claude Code you can choose which commands need approval. Keep deployment, deletion and package installs on “always ask”.
  4. Keep secrets out of reach. Store API keys in environment files the agent does not print, as covered in how to store AI API keys safely. WorldofAI notes that Anthropic says keys in Claude Code’s new cloud environments are stored securely and not visible to Claude, but it is still good practice.
  5. Write a clear brief. One goal, what is out of bounds, and “stop and ask if you need to change anything outside this folder”.
  6. Turn on notifications only for approvals, so you are not checking the phone every minute.

💡 Tip: Add one line to your brief: “Before asking for approval, summarise in two sentences what the change does and how to undo it.” It makes phone reviews much faster.

How do you review a change on your phone, step by step?

  1. Read the agent’s summary first. What is it trying to do, and why now?
  2. Check the file names. Are they the files you expected? A surprise file is a reason to pause.
  3. Check the size. A few lines is fine on a phone. Hundreds of lines can wait.
  4. Look for red-flag words in commands: delete, remove, drop, force, deploy, publish, install, sudo.
  5. Ask if unsure. Reply “Explain this change and how to undo it” rather than approving.
  6. Approve or reject. If you reject, say why, so the agent tries a better route.
  7. Review properly later. Look at the full branch on your computer before merging anything.

Approve, ask or wait? A quick decision table

Request from the agentWhat to do on your phone
Edit text or styles on one pageApprove
Add or run testsApprove
Install a new packageAsk what it is and why, or wait
Delete files or foldersWait until you are at your desk
Change database or payment settingsWait, and review in full
Deploy or publish to your live siteWait, and review the whole branch first
Read a file that holds secretsReject

What mistakes should you avoid?

  • Approving out of habit. Repeated prompts can lead to tapping “yes” without reading.
  • Giving blanket permission to save taps. It removes the safety net you rely on.
  • Working on the live branch. One bad approval can break your site straight away.
  • Ignoring hidden instructions. Agents can pick up instructions from files or web pages. Our guide on stopping an AI coding agent following hidden instructions explains what to watch for.
  • Skipping the final review. Phone approvals are steps, not sign-off.

🔎 Unconfirmed / rumour: WorldofAI also reports that Anthropic appears to be working on a collaborative mode for Claude Code projects, which would let several people share context on one project. It has not been released, so approval rules for teams may change if it arrives.

Frequently asked questions

Can I approve Claude Code changes from my phone?

According to WorldofAI, Claude Duet now supports Claude Code sessions, letting you monitor, approve changes and send instructions from your phone. Check your plan and app version for availability.

Is it safe to approve AI code changes on a phone?

Yes, for small and reversible changes on a separate branch. Leave deletions, deployments, database and payment changes for a full review at your desk.

Does the session keep running on my computer?

It can run locally on your computer or in the cloud, as described in the video. Your phone acts as the remote control.

What if I approve something by mistake?

If you work in Git on a branch, you can roll the change back. That is why version control comes first.

Should I give the agent permission to run any command?

No. Keep risky commands such as installs, deletes and deploys on “always ask”, even if it means more prompts.

The bottom line

Approving agent changes from your phone is a real time saver, as long as you keep phone approvals for small, reversible steps. Set up Git, a working branch, sensible permission limits and secrets the agent cannot see before you leave your desk. Then use your phone to keep things moving, and save the big decisions for a full screen. Your first step: add the two-sentence summary rule to your next agent brief.

author avatar
Garry Knight
I'm Garry Knight, the person behind Prodify Digital. I write about email list building, email marketing, SEO, AI search and the tools that connect them. My aim is to make online marketing easier to understand, so creators and small business owners can make informed decisions about building an audience and keeping people engaged. Here you'll find straightforward guides and product reviews that explain what something does, where it fits and which limitations matter. The focus is on clear explanations and useful next steps—not hype, shortcuts or promises of easy earnings.

Leave a Reply