Stop babysitting your AI coding agents

Watching the terminal to catch when an agent finishes or needs a decision is a bad use of attention. Here is how to break the habit on macOS.

By Piotr Jura, creator of AgentNotch. Tested on macOS with AgentNotch v2.2.4.

❯ claude

● Reading docs/licensing.md

⎿ Read 240 lines

⎿ Grepped "claimToken" (11 files)

● Wiring the license claim endpoint

⎿ Edited lib/license/store.ts (+38 −6)

⎿ Edited app/api/license/claim/route.ts (+21 −3)

⎿ Ran npm test (42 passed)

● Run npm run db:migrate?

❯ 1. Yes 2. No, tell Claude what to do

waiting for your approval

5h34%2h11mWeek41%~4dWeek58%~3dMonth22%~13d

You know the posture. The agent is working, you've nothing to do until it stops, and you watch the terminal anyway because looking away means missing the moment it needs you. So you sit reading output you won't act on.

It feels responsible. It's mostly wasted attention. An agent needs you at the pauses: when it finishes, stalls, or stops for approval. A terminal bell or a Claude Code Stop hook helps, but each fires once and never says which of several agents is calling.

Where AgentNotch fits

AgentNotch parks that state in the menu bar. It watches your Claude Code, Cursor, Codex, Kimi and OpenCode sessions, plus the Cowork sessions in your Claude app, flags the one waiting on approval, stalled, or done, and shows context and plan usage per session. Click one to land in its window. You get pulled back only when it matters. Nothing leaves your Mac: it never reads your code or keys.

Questions

How do I stop babysitting my AI coding agents?: Replace the habit of watching the terminal with a signal that reaches you when an agent actually needs something. Turn on notification hooks or a terminal bell for a single agent, and use a menu-bar monitor like AgentNotch to flag the one session that needs you when you run several.
Why do I keep watching the terminal while my agent works?: Because the agent gives no reliable cue when it stops, so watching feels like the only way not to miss the moment it needs you. Once you have a signal that surfaces that moment on its own, the watching becomes unnecessary.
Is it a waste of time to watch a coding agent run?: Mostly, yes, since a coding agent only needs you at the pauses, when it finishes, stalls, or needs approval. The stretches between need nothing from you, so watching them is time you could spend elsewhere.
What is the real cost of babysitting agents?: The cost is the work you don't do while you sit and watch, plus the context-switching tax of checking a window every few minutes. Both add up faster with each extra agent you run.

Keep reading

See which agent needs you.

Try AgentNotch free for 14 days, no card. Watch Claude Code, Cursor, and Codex from your Mac's menu bar.

Start free trial See pricing

Updated 2026-07-17

Where AgentNotch fits

Questions

How do I stop babysitting my AI coding agents?

Replace the habit of watching the terminal with a signal that reaches you when an agent actually needs something. Turn on notification hooks or a terminal bell for a single agent, and use a menu-bar monitor like AgentNotch to flag the one session that needs you when you run several.

Why do I keep watching the terminal while my agent works?

Because the agent gives no reliable cue when it stops, so watching feels like the only way not to miss the moment it needs you. Once you have a signal that surfaces that moment on its own, the watching becomes unnecessary.

Is it a waste of time to watch a coding agent run?

Mostly, yes, since a coding agent only needs you at the pauses, when it finishes, stalls, or needs approval. The stretches between need nothing from you, so watching them is time you could spend elsewhere.

What is the real cost of babysitting agents?

The cost is the work you don't do while you sit and watch, plus the context-switching tax of checking a window every few minutes. Both add up faster with each extra agent you run.