Stop Babysitting Your AI Agent: An Async Coding Workflow

Claude Code can work for twenty minutes on its own, but most people still sit and watch it, because the moment they leave it stops to ask a question. That turns a fast agent into a slow one. This post lays out an async workflow that lets you walk away: how to start tasks so they run longer without you, how to get pinged only when it matters, how to answer in seconds, and where to draw the line on what you approve from a phone.

The bottleneck is you, not the model

Watch a typical agent session and count where the time goes. The model reads files, writes code and runs tests quickly. Then it asks for permission to run a command, and the session sits idle until you notice. Then it finishes, and the result sits unreviewed until you notice that too.

The waiting is the slow part, and it's all on the human side. The fix isn't to watch harder. It's to be reachable without watching: let the agent run, and get a signal only when your input changes something.

Andrew Leech's write-up on using Claude Code from anywhere is a good example of the do-it-yourself version: Tailscale for networking, Eternal Terminal for connections that survive sleep, tmux to keep sessions alive, and a phone as a thin client. As he puts it, "I can start something on my desktop, walk away, and check on it from my phone." The workflow below gets you to the same place with less setup.

Start the task, then walk away

A session that runs well unattended starts with a prompt that answers the questions Claude would otherwise stop to ask. Before you leave, spend one more minute on the prompt:

• Scope. Which files or folders it should touch, and which it shouldn't.
• Done. What finished looks like, in a way Claude can check itself.
• Verification. The exact command that proves it works.
• Defaults. What to do when it hits a choice you don't care about.

For example:

Add rate limiting to the /api/upload route.
Only touch src/server/ and its tests.
Done = the new tests pass and npm test is green.
If a library choice comes up, use what's already in package.json.
If something outside src/server/ seems to need changing, stop and ask.

That last line matters. You want Claude to stop for the questions that need you and decide everything else.

You don't have to be at the desk to start. With Coder IDE you can start a new Claude Code session in any project on any of your connected computers from your phone or browser, and the computer picks it up even while you're away. Claude Code keeps working when your phone sleeps, you switch apps, or your connection drops.

Get pinged only when it matters

Notifications are what make walking away safe. Coder IDE sends a push notification when a Claude session needs your approval, is waiting for your reply, or stops unexpectedly. Everything else, like a session that's busy or one that finished cleanly, waits in a list grouped by what needs you, so you can check it when you choose.

That list covers every session on every connected computer, including ones you started in a terminal, in VS Code, or in the background. If you run several sessions at once, managing multiple Claude Code sessions from your phone covers that side.

Anthropic's own Remote Control also sends push notifications for permission prompts and questions, and it may be all you need if you work in one session on one machine. Use whichever fits how you work.

Decide in ten seconds

When the notification arrives, the goal is a fast, correct decision. Coder IDE shows you the full request before you approve or deny it, from your phone or a browser. A simple rule set keeps those decisions quick:

For replies, short and specific beats long and thoughtful. "Use option B, keep the old function as a wrapper" unblocks a session. A paragraph of context you could have put in the original prompt usually means the prompt needs work next time.

If a request surprises you, deny it and reply with what you expected instead. A denied request with a one-line reason costs Claude less time than an approved mistake.

Review on the phone, merge at the desk

When a session finishes, you can do most of the review away from the desk. Coder IDE shows a session's changes file by file. Comment on any line, then send all your comments back to Claude in one message, and it gets to work on the next round while you do something else.

When a fix is faster by hand, Open in VS Code takes you to the exact file and line in the real editor running on your computer. If the change has a UI, start the dev server, turn on Allow preview for its port in the Ports panel, and open the private link to see the result beside your code or in its own tab.

Keep the final merge at the desk. A phone is fine for reading a diff; it's a poor place to judge how a change fits the rest of the codebase. Reviewing Claude Code's diffs on your phone goes into what to check on a small screen and what to defer.

Voice for hands-busy moments

Sometimes you can't type: you're carrying groceries, cooking, or walking the dog. The Android app's Voice tab lets you hold the mic, say what you want, and hear the answer read back. It's well suited to short, low-risk instructions:

• "Ask Claude what's left on the upload task."
• "Run the tests again."
• "Tell Claude to keep going with the plan."

Save anything that needs you to read code for when you can look at the screen.

Guardrails: never approve what you can't see

Async work goes wrong when convenience beats attention. A few rules keep it safe:

• Read the request before you tap. If the notification doesn't show enough, open it. Coder IDE shows the full request for exactly this reason.
• Work on a branch or worktree. Anything the agent does is then one git reset away from gone.
• Be careful with skipping permissions. Claude Code can run with permissions bypassed, and some people do that inside git-tracked projects to avoid prompts entirely. It removes the bottleneck and the safety net together. If you use it, use it for well-scoped tasks in a project you can roll back.
• Keep secrets and production out of reach. Don't give an unattended session credentials it doesn't need.
• Defer what you can't judge. "I'll look at this at my desk" is a valid answer.

Wrap-up

An async workflow is mostly about being reachable instead of present: a prompt that anticipates questions, notifications only for the moments that need you, quick rules for approvals, and reviews that happen wherever you are. Coder IDE puts those pieces on your phone and in your browser while Claude Code keeps running on your own computer.

Start free. The free plan covers one computer, which is enough to try this workflow end to end.