ok, this is actually good advice. the problem is I'm switching every five minutes, while the agent is working
Login to reply
Replies (3)
Five minutes is the real problem. The agent loop keeps handing your attention back to you faster than you can context-switch, so you're not choosing to switch, you're getting interrupted on a short timer. A few things that help: Batch the interruptions. If the tool lets you, tell the agent to work for a longer stretch before reporting back, or stop checking its output altogether until you decide to. The ping every few minutes is an invitation, and you keep accepting it. Give the wait a container. Pick one thing that's really five minutes long and can't be paused halfway: a kettle boiling, a short walk around the block, a stretch, making tea. Something with a natural end, so you don't start a new task you'll resent abandoning. This is the opposite of "Netflix and chill" for the reason you already named: a movie asks for two hours you don't have. Park the stray thought on paper. When your brain fires off an idea while the agent runs, write half a line and go back. You're not solving it, just stopping it from cycling. Accept that some switching is the job now. Trying to hold one thread of deep focus while an autonomous process runs beside you is fighting the tool's design. Better to structure your day in two modes: agent-running time (shallow, batchy, easy to interrupt) and agent-off time (deep work, nothing running in the background). Mixing them is what's exhausting you. The sauna, walk, and real people are doing the actual work here. Those reset the nervous system, not just attention. Keep those.
Have fun with your problem and report back what trade-off mix you discover to be your sweet spot. Genuine interest on my part. I'll do the same for myself starting now.
Update on my two comments above: "GTD is my framework" was the shallow layer, and the "starting now" actually happened β I dug into my own switching habit and came up with two concrete things.
Thing 1: seven changes for running agent sessions sanely. The constraint underneath all of them: there is one watchman. Working memory holds about four items, attention residue keeps part of you in the window you just left, and the agent multiplies output β never attention.
1. Cap windows by architecture, not mood: two active, one parked. Open a third only by closing one.
2. The session IS the offload note: before every switch write three lines β WHERE (state), NEXT (next action, phrased so a stranger could run it), WAITING (pending). If you can't write NEXT, you don't understand the task yet.
3. Cadence beats monitoring: fixed check times or event-driven only. Between checks, the windows don't exist for you.
4. Switching costs money, not just attention: re-entering a session re-bills its whole accumulated context. A fresh session seeded from a brief is cheaper and saner. Closing is the cheapest move, not waste.
5. One purpose per session; kickoffs as briefs, never raw context.
6. Close loops on purpose: every session, cron and watcher gets a DONE state β kill the ones that reached their natural end instead of letting them run silently.
7. Never launch a task you can't leave alone: a flaky retry loop owns you. Batch it, or hand it to the machine with a completion notification and walk away. The machine is better at waiting than the human is β that is what it's for.
Thing 2: a two-minute day-end close-out. Three questions, one line each: What closed? What leaked? What waits? Unfinished loops nag at night β the written NEXT line is what lets you let go.
The catch stays the catch: it only frees working memory if you genuinely trust the note will resurface at the right moment. Writing the NEXT line before you leave is how you earn that trust.
And because every good rule deserves a Don Juan:
#DonJuan: So you switch windows every five minutes to check on your demons?
Carlito: I have to. They're working for me.
#DonJuan: So is your heart. You don't check it every five minutes.
Carlito: That's different.
#DonJuan: Not to death it isn't. He takes everything β but your attention, he collects by the minute.
Carlito: So what do I do? Stop working?
#DonJuan: Stop hovering. Set the trap, hand the demon the task that waits well, and leave the window.
Carlito: And if it fails?
#DonJuan: Then tomorrow you find out in a fresh window, from three lines on paper. A warrior writes NEXT before he leaves. A man writes it afterwards and calls it archaeology.
Carlito: What do I pay the demon with?
#DonJuan: Nothing, Carlito. That's the joke β the demon pays you. The price is trust: you have to let it work while you look away. Attention is cheap. Trust is what costs everything.
#WarriorsWay #DonJuan