The work you can do while waiting (without paying the switch tax)

In the last piece we landed on an uncomfortable fact: the worst thing to do while an AI agent works is jump into deep work on another project. The resumption cost is brutal, you stop catching the agent’s mistakes, and the measured result is that you go slower while feeling faster.

But “do nothing” is not an answer either. You have a real gap, several times an hour, and staring at a progress bar is not a plan. So the honest question is narrower and more useful: what work actually fills a gap cheaply?

High-residue versus low-residue work

The cost of a switch is not fixed. It depends on how much mental context you have to tear down and rebuild. That gives you two very different categories.

High-residue work drops you into a different mental model. Opening a second project’s codebase. Picking up a design problem you set down yesterday. Anything where you have to reload an entire context before you can think. This is the expensive switch. It carries the full resumption lag, and it leaves residue that bleeds back into the original task when you return.

Low-residue work stays in roughly the same headspace, or needs almost no headspace at all. It is adjacent, shallow, and ready to grab:

• Reviewing what just came back. Reading the diff the agent produced. You are already in that context.
• Teeing up the next instruction. Writing the spec for the next task while the current one runs. This is the highest-leverage gap-filler there is, because a sharper spec makes the next agent run longer without you.
• Knocking out a genuine quick win. A two-minute task that is fully specified in your head already. No reload required.
• Light triage. Sorting what is ready, deciding what is next, clearing small decisions.

The first category costs you the rest of your afternoon in small pieces. The second category is nearly free, and some of it (writing the next spec) actively makes your agents more autonomous. The whole game is to fill gaps with the second kind and refuse the first.

The problem is that low-residue work is hard to find on demand

Here is the catch. In the moment, when the agent is running and you have eight minutes, the low-residue options do not present themselves. Your eye lands on the other project’s open tab, which is exactly the high-residue trap. Finding the cheap thing to do requires scanning everything you have going, judging what is ready and how big it is, and picking, which is itself a small context switch.

So the right move is to have that decision already made for you.

On Deck

That is what we built. On Deck is a single view that answers “what should I do in this gap” across everything you are running.

• It pulls one next move from each active project, so your whole board is in front of you at once.
• It ranks by switching cost. The project you most recently worked is pinned at the top and labeled “stay in flow,” because that is your lowest-residue option. The rest follow, warmest first.
• It has a gap-size filter. Tell it you have a few minutes and it surfaces the quick wins. Tell it you have a real block and it shows the deep work. The size of your wait decides what you see.
• You act inline. Start a focus session or mark something done right from the queue, so grabbing a move does not require navigating into the project and paying the reload cost just to get started.

It is deliberately not a full project view. A full view invites you to dive in, which is the trap. On Deck is the opposite: a thin, ranked surface that hands you the cheapest useful thing and gets out of the way.

Why this is the right shape

Recall the fan-out math: the number of agents you can run equals their neglect time divided by your interaction time. The two ways to run more without drowning are to raise neglect time (agents that run longer alone) and lower interaction time (faster to direct and review).

On Deck is built to push both. When your default gap-filler is “write the next spec” or “review the last diff” instead of “dive into project B,” your specs get sharper (neglect time goes up) and your reviews stay current (interaction time goes down). The gap stops being a place where your attention fragments and becomes the place where you set up the next hour to run itself.

The waiting is not going away. The agents are only going to get faster at handing the gap back to you. The skill that matters now is what you do with it.