← Writing

The category of AI coding tools just split: tools that let you walk away from the screen, and tools that still need you sitting there.

You removed the friction from the workflow. The last friction is the desk.

You stopped babysitting (post #4). You went parallel with worktrees (post #5). But you’re still sitting there. Watching terminals. Waiting for the ping. Tethered to the screen.

The AI doesn’t need you at the keyboard anymore. The question is whether your tool knows that.

The Split

The category of AI coding tools just divided into two tiers.

Tier 1 — Tools that let you walk away:

  • Telegram/Discord channels — your agent pings you when it’s done. You reply with the next task from your phone.
  • Remote control — pick up your session from a browser, tablet, or phone. No terminal required.
  • Scheduled agents — cron-triggered tasks that run in the cloud while you sleep. Wake up to completed work.
  • Slack/GitHub triggers — an event fires (new issue, failed test, PR review request), an agent spins up, a PR lands.
  • Cloud VMs — agents running on their own machines, shipping PRs autonomously.

Tier 2 — Tools that still need you in the chair:

  • Copilot CLI — no native remote session. Community bridges exist, but you’re duct-taping it together. You have to be at the machine.
  • Cline — no remote session, no push notifications. The agent works when you’re watching.

Why This Matters

The friction progression over the past year:

  1. You stopped approving safe commands (deny list, not allow list)
  2. You stopped watching every step (trust + hooks)
  3. You stopped managing context manually (auto memory)
  4. You stopped working sequentially (worktrees)
  5. You stop being tethered to the desk ← we’re here now

Each step compounded. The cumulative effect is an order-of-magnitude difference in what one developer can get done in a day.

If your tool still requires your physical presence at a machine, you’ve optimized every other variable and left the biggest one in place.

The Practical Setup

For async operation with Claude Code:

# Start a task, get a Telegram notification when done
claude --remote --notify telegram "Implement the rate limiting middleware, 
add tests, update the API docs"

You start the task from wherever you are. The agent works. You get a ping. You review the diff from your phone, approve or request changes, and move on.

The session persists. The work doesn’t stop when you’re not watching.

The Right Question

The question isn’t “which AI tool has the best model?” Current frontier models are close enough on benchmarks that the answer changes month to month.

The right question is: after you’ve set up the task, where do you have to be?

If the answer is “at my desk, watching a terminal” — you haven’t removed all the friction yet.

AI Minus the Friction #6. The complete series: Read all 6 posts →