What to Do While Claude Code or Codex Works: A Simple Protocol for the Two-to-Twenty-Minute Gap in Every Agent Run

You hand Claude Code a refactor, or ask Codex to run the full test suite and fix what breaks, and the terminal starts scrolling. The next five to twenty minutes are the strangest part of working with a coding agent: you are not needed, you cannot really start something else because it might ask a question, and watching the spinner is both pointless and weirdly compelling. Most developers fill the gap with Slack, a news tab, or a second agent, and arrive at the end of the day having sat completely still for eight hours.

This post is a protocol for those minutes: a decision rule for what to do based on how long the run will take, a set of movements that fit in the gap without changing clothes or leaving the desk, and the mechanics of doing it with NotchFit, which was built for exactly this situation. It also covers when not to work out during a run, because an agent that needs you every ninety seconds is a different problem from one that needs you in fifteen minutes.

Step one: estimate the gap

The useful question is not "what should I do while Claude works?" but "how long is this run, and will it interrupt me?" Three rough buckets cover most tasks:

• Under two minutes (a small edit, a single file, a quick question): stay. Stand up if you have been sitting a while, look at something far away, roll your shoulders. Do not start a set; you will be interrupted mid-rep and the whole thing becomes annoying.
• Two to ten minutes (a multi-file change, a test run, a migration): one set. This is the sweet spot for a micro-workout and the gap NotchFit is designed around.
• Ten minutes or more (a large refactor, a long build, an agent working through a list): two or three sets with rest between them, or leave the desk entirely and let the notification bring you back.

You will get better at estimating with practice, and the agent helps: when a task starts, NotchFit's notch panel shows the AI's live progress on the right-hand side, so you can see whether it is still exploring the codebase (long) or already writing the summary (short) before you commit to a set.

Step two: the set that fits the gap

A micro-workout during an agent run has constraints a gym session does not. It has to be silent enough for an office, need no equipment, not make you sweat through a shirt, and be stoppable the moment a permission prompt appears. NotchFit's exercise library (239 animated moves, 198 of them photo animations of real athletes) is built for that, and the way the app runs a set is worth understanding before you start:

• Sets start only when you tap Start set. A task starting does not drop you into a workout; the notch opens with a Start workout card and the agent's updates feed, and you decide.
• Each set has a rep or time countdown ring with Done, Skip and Pause. Skip is a real button, not a failure state: if the agent finishes early, skip the rest and go.
• Rest timers run between sets, so a ten-minute run naturally becomes two sets and a rest rather than one long effort.
• Done and permission notifications that arrive mid-set wait until the set ends, so you are never interrupted mid-rep. We explained the reasoning in why mid-set notifications wait.

Which exercises? That depends on you, and the honest answer is that a plan written for your body beats any generic list. NotchFit's weekly plan is written by your own agent: run /notchfit:plan in Claude Code or $notchfit-plan in Codex, answer its questions about age, fitness level, goal, equipment, days per week, minutes per day and injuries, and it produces a weekly plan using only exercises the notch can animate. There is also an offline quick-plan generator and a starter plan by default, so you can begin today and refine later.

Step three: decide how far you will go from the keyboard

For short and medium runs, stay at the desk: the point of the notch panel is that the agent's progress and your exercise are in the same glance, at the top of the screen, and the moment the agent needs permission you are one tap away. For long runs, you have two options and should pick deliberately.

Stay in the room if the task is the kind that asks questions: anything touching files outside the repo, anything that installs packages, anything in a new codebase. A permission prompt that waits ten minutes for you is ten minutes of wasted agent time, and NotchFit's notification only helps if you are within earshot or eyeshot. Optional spoken alerts, using on-device macOS voices, are useful here if you are across the room; there is a mute button in the notch for when you are not alone.

Leave the desk if the task is self-contained and pre-approved: a test suite, a build, a documented refactor with permissions already granted. Then the right move is not a set at all; it is a walk, water, or a window. The notch will show the summary when you return. If you want the agent to carry on without any prompts at all while you are gone, the background-session and effort settings in Claude Code matter more than anything NotchFit does; we covered them in the background sessions guide.

The anti-protocol: when not to do this

• Interactive sessions. If you are pairing with the agent, approving every tool call, or iterating on a prompt, the gaps are seconds, not minutes. Use the Hide workout toggle and let the notch be a pure Claude Code and Codex status panel for the afternoon.
• Right before a meeting. One set is fine. A set you cannot finish because the call starts is a set that leaves you slightly out of breath on camera.
• Shared screens. The notch panel is at the top of your display. If you are screen-sharing your whole screen, the animated exercise is visible to everyone on the call. Hide the workout or share a window instead.
• Pain, not discomfort. NotchFit is a workout guide, not medical advice. The plan asks about injuries so it can avoid movements that aggravate them; if something hurts, skip it and tell the plan next time.

Making it stick

The reason this works where "take a break every hour" apps fail is the trigger. A timer interrupts you at an arbitrary moment, usually mid-thought, and you dismiss it. An agent run interrupts nothing: the agent has just taken the work off your hands, and the decision is only what to do with the minutes it gave back. NotchFit attaches the habit to that moment, keeps a streak and a history heatmap, and throws confetti when you finish a set or a workout, which is sillier than it sounds and more effective than it should be. The longer argument for why the trigger matters is in building a developer movement habit with streaks.

A realistic day looks like this: six or eight agent runs long enough for a set, each set one to three minutes, a rest timer here and there. That is ten to fifteen minutes of movement you would otherwise have spent reading the spinner, spread through the day in the gaps the agent created. Nobody will mistake it for training. It is also, for a lot of developers, more than they moved yesterday.