AI Agents and the Art of Doing Nothing

The "agentic workflow" is essentially just a fancy way of saying I now spend 40% of my workday staring at a loading spinner while some LLM agent pretends to be a senior engineer. We've traded the pain of writing boilerplate for the existential dread of watching a terminal window execute npm install for the fifth time because the agent forgot it already did it.

It's a weird psychological trap. When I was coding manually, I was in the flow. Now, I'm just a glorified babysitter for a bot. The second the agent starts a "deep dive" into my file structure, my brain instinctively screams "OPEN REDDIT" or "CHECK YOUTUBE." It's the ultimate productivity paradox: the more the AI does, the more I struggle to stay focused.

If you're trying to survive this without losing your mind to a 3-hour rabbit hole of cat videos, here is how I'm attempting to handle the downtime:

The "Parallel Tasking" Strategy

Instead of treating the agent like a synchronous process, I've had to start treating it like a junior dev who takes an hour to reply to a Slack message.

1. Queue-based working: I stop expecting immediate results. I give the agent a massive task, then I physically move to a different project or a different branch.
2. The "Analog Buffer": I keep a physical notepad next to my keyboard. While the agent is "thinking" or executing a loop, I sketch out the next three prompts or map out the logic. It keeps the brain in "dev mode" without the temptation of a browser tab.
3. Strict Time-Boxing: If the agent hasn't solved the bug in three loops, I kill the process. Nothing kills focus faster than watching an agent hallucinate in a circle for ten minutes.

The real goal of a real-world AI workflow shouldn't be "how do I wait better," but how to structure tasks so the agent isn't the bottleneck. If you're spending your whole day waiting for a CLI tool to finish, you're not an engineer anymore—you're just a prompt operator with a lot of free time.