Your AI agent is waiting for you. Your attention isn't. How much time do you lose to this?
by•
Daily reality: Claude Code in one pane, Codex in another, Cursor in a third, plus a few shells. An agent stops and waits for "yes". By the time I find the right pane, 20 min gone.
Just me? What's your setup — notifications, layout tricks, or you've accepted the tax?
149 views
Replies
Same experience here . The AI isn't usually the bottleneck finding the right terminal and remembering what each agent was doing is.
@jacob_spencer2 Exactly — the bottleneck moved. The model is fast enough now that the friction is on our side: locating the pane, then reloading "what was this one doing again?" in your head.
That second part is why PokePoke's notification carries the last message from the agent, not just "waiting" — so you get the context back before you even switch to the pane.
Honestly this is exactly why I built one for myself — PokePoke. Menubar app, pings me when any agent (Claude Code / Codex / Cursor) is waiting, click the notification and it jumps straight to that pane.
Curious if others solved this differently — what am I missing?
probably 15-20 min a day easily, but the actual cost isn't the wait, it's that I forget which of the 3 panes was even blocked and end up re-reading the last few messages in all of them to remember. a simple terminal bell or OS notification the moment it's actually waiting on you (not just running) would fix most of it, but half these tools don't distinguish 'thinking' from 'blocked on approval' in any visible way.
@omri_ben_shoham1 Yeah, the "thinking vs blocked on approval" distinction is the hard part. A bell that fires on every pause becomes noise fast; one that only fires on real blocks needs to actually understand what the agent is doing.
That's basically why I ended up building PokePoke — it hooks into Claude Code / Codex / Cursor at the approval moment specifically, and clicking the notification jumps you to that pane so you skip the "which one was it" step you described.
Buffup.AI
This is painfully real 😂 By the time I find which pane is waiting for approval, I have already lost 20 minutes. I need one universal agent needs attention notification.
@sansa_grey Same 😂 The "universal" part is what got me too — I didn't want a different setup for Claude Code vs Codex vs Cursor. Ended up building PokePoke for exactly this: one menubar notification across all of them, click to jump to the pane.
Would love your take if you try it.
My biggest issue is having 5 agents running and no idea which one just stopped. We need a central dashboard for agent status.
@kat_pim Honestly I'd push back a little — a dashboard sounds right but adds another surface to keep watching. What actually fixed it for me was inverting it: no dashboard, just a strong notification the moment one stops, that also takes you to the exact pane on click.
That's what PokePoke does. Been running 5+ agents daily without a dashboard since.
35 minutes later: Oh… that agent was waiting for me.” 😭 I now keep one dedicated pane for anything that needs human approval.
@ashir_murtaza1 That works until you can't predict which task will need approval — half the time I think "this one's autonomous" and it stops on the third tool call anyway. Also the discipline itself has a cost: one more rule to remember while you're deep in something else.
That's why I went the other way with PokePoke — no dedicated pane, no rules, just let the agent tell you when it needs you.
I keep missing approval prompts because they diappear behind other windows. A small unified notification panel across every coding tool would save me time. Have you tried any workflow that solves this without adding more complexity?
@eoin_bishop Yeah the "behind other windows" part is why I ended up putting it in the menubar instead of a regular window — it can't get buried. Same reason no dashboard, no extra app to open: notification shows up, click, you're in the pane.
Built PokePoke for exactly this. Works across Claude Code / Codex / Cursor in one place.
me using multiple agents at once sounded productive until context switching became the biggest delay. I now limit active sessions. Has anyone measured whether fewer parallel agents actually finish work faster?
@gaspard_dupuich Interesting angle. I think the answer depends on how fast the LLM itself is — if inference gets 3x faster tomorrow, the "context switching tax" you're describing shrinks a lot.
But honestly, the thing that changed for me isn't measured productivity. It's that once agents were a thing, I can't stop myself from spinning up tasks in idle moments — not even work stuff. "Can my terminal look prettier?" "What if I try that CLI I bookmarked?" Parallel agents feel less like a productivity strategy and more like… a new default behavior.
Not sure if that's better or worse, just different.
The future should be agents handle the routine stuff, humans only get pulled in for ambiguity. Right now, we are still manually babysitting too many workflows.
@ashley_james Agreed, and I think we'll get there — but not tomorrow. Model reliability, cost, and trust all need to move together, and any one of them lagging means we're still babysitting. Looking forward to that day, but building for the current reality in the meantime.
I think this gets to the real bottleneck. We've become much faster at generating work, but switching attention between agents and remembering why each one is doing something still feels expensive. I'm curious whether the answer is better interfaces or a shared context layer that keeps every agent aligned with the original intent.
@nextmark This is the right framing, honestly. And I'd argue both directions are already being explored — just by different products.
The "shared context layer" you're describing is basically what tools like cmux and herdr are building — agent orchestration, shared memory, keeping parallel agents aligned. That direction reduces the volume of interruptions by making agents smarter about when to ask.
The other direction — better interface for the moment an agent does pull you back in — is more under-invested. Even with perfect orchestration, agents will still hit ambiguous moments and need you. When that happens, "which agent, which pane, what was it doing" is still a manual cost. That's the gap I've been working on.
Both matter. Just curious which one hurts more for you day-to-day.