Running most of my work through agents has moved the bottleneck twice. First, writing stopped being the constraint and reviewing became it; I wrote about the furniture we built for that in The Desk. Then reviewing documents got fast, and the constraint moved again, to something smaller and more frequent: the judgment call.
Agents hit them constantly. Should the invoice show the fee breakdown or roll it up? Does this workflow belong to the platform or the product? Two features both claim the same field; which one is right? None of these are things an agent should decide alone, and all of them arrive at the worst possible moment: mid-build, mid-session, while I'm doing something else.
The naive mode is interrupt-driven. Every agent that hits a question pings me, and I context-switch into a decision I have no background loaded for, answer from a cold start, and go back to what I was doing. Multiply by a dozen sessions a week and the interruptions aren't the worst part. The worst part is the quality of the answers. A judgment call answered in the gap between two other thoughts is a coin flip wearing my name.
So we stopped doing that. The pattern that replaced it has two halves: save the questions up, then walk through them in a specific format at a moment I choose. Around here it's called the gate walk, because each open question is a gate that some piece of work is stopped behind.
One honest query
The saving-up half sounds trivial and isn't. For a queue to be worth sitting down with, it has to be complete, and completeness is a discipline, not a feature.
Every question waiting on me carries the same tag in the tracker. That's the whole mechanism. What makes it work is the rule wrapped around it: the tag query has to be the honest answer to "what needs me." Any session that hits a judgment call files it and tags it, right then, with the evidence attached. If a question shows up written in prose somewhere (a handoff note, a status summary) without the tag, that's drift, and whichever session finds it fixes it on the spot. A question that lives only in prose is a question that will be asked twice or lost entirely.
There's a subtler failure the rule has to catch. Sessions like to mark things "parked: waiting on Aaron, no rush," and a parked label is a claim, not a fact. If the underlying note actually contains open questions and nothing anchoring the park (no ruling from me, no standing instruction), it's a live gate someone talked themselves out of surfacing. We learned to check. The queue only stays trustworthy if "parked" has to show its receipts.
The payoff for the discipline is that "what needs me?" becomes a query I can run any time, and the answer is neither a guess nor a scavenger hunt. Some weeks it's two items. Some weeks it's eleven. Either way, nothing is waiting on me that I can't see.
The walk
When I sit down, the queue arrives in a deliberate order: things only my hands can do first (a token refresh, a publish that needs my account), one-word approvals next, and the deep judgment calls last, sorted by how much work each one unblocks. Parked items get listed at the bottom for completeness and explicitly not asked about. Then we go one gate at a time.
Each gate arrives as a package with four parts.
Background. What exists, what's already settled, and what's genuinely open. Evidence-anchored, but in plain English, and always with a concrete scenario, because abstractions hide the question. "How should site attribution work?" is fog. "The Jacksonville office books a trip on a Teterboro-based aircraft: whose board does it show up on?" is a question I can actually answer.
One recommendation. Committed, singular, no menu. A list of options with no opinion is the agent outsourcing its homework to me. If it has read all the evidence and can't form a lean, the package isn't ready.
Rationale. Why the recommendation, anchored in decisions already on the books where they exist, with the honest counter-argument stated instead of buried.
Implication. What follows from each branch, not just the recommended one. If I pick the other path, this section is where the real bill is itemized: what gets rebuilt, what gets slower, what other decision it collides with.
Then: your call. And it waits.
The format sounds heavy. The effect is the opposite. Because the package did the loading, my answers come out terse: "Agreed." "Yes, but only for closed trips." "C, but automatically as prep." A one-word answer to a well-built gate is not a careless answer; it's what a decision looks like when the deciding was made easy and only the judgment was left. The terseness is the proof the packaging works.
Recorded before the next gate
Here is the half I'd defend hardest: nothing about the next gate happens until the previous answer is a record.
Each ruling gets written down as a decision, and the decision quotes my exact words, dated, with the context they were said in. Not a paraphrase. Paraphrase is where drift starts: a summarizer smooths a condition off the edge of an answer, and three sessions later the condition is gone. When the answer was "C, but automatically as prep," the "but automatically as prep" clause is the ruling, and it goes into the record verbatim.
The record isn't just prose either. If the ruling implies work, that work gets filed as tracked items at decision time, linked to the decision, so the executable half of an answer can't evaporate. If the answer was about ordering ("land X before Y"), it becomes a dependency edge in the tracker, not a sentence someone has to remember. And the moment a gate is resolved, the tag comes off, and whatever was stopped behind it dispatches. Agents start building on ruling three while I'm still reading gate four. The walk and the work run concurrently.
Then the next gate opens, and it opens by naming the record the previous answer became. Said once, recorded immediately, binding on every future session. I never repeat myself, I never police whether something I said got captured, and six months later "why does it work this way" has an answer with my name and a date on it.
Listening rules
The part I didn't expect to matter most is a set of rules about listening, each one learned by getting it wrong.
The big one: when my answer addresses a different question than the one asked, the answer is still real. Early on, an agent asked me how sites should attribute trips, and I answered with how aircraft positioning should work, because that's what the scenario made me think about. The wrong move is to bulldoze my answer into the question's shape and record a ruling I never made. The right move, and what the format now requires, is to record what I actually said as its own decision about the thing I was actually addressing, then re-pose the original question with a sharper scenario. That session produced two decisions where a sloppier listener would have produced one wrong one.
The rest of the rules are smaller but the same shape. When I say the question doesn't make sense, the compression was the bug, and the fix is a more concrete re-pose, not a defense of the original wording. When an agent infers anything beyond my literal words, the inference gets marked as an inference, with a standing offer to amend if it over-read. And when a gate is a default I could veto and I say nothing, the default stands and nobody asks again. Silence is an answer too; nagging is a tax on the whole system.
The invariant
The gate walk ends with a test you can run. After the last gate, the tag query should return only items that are parked on purpose, with their parking anchored to something I actually said. If anything else is still in the queue, the walk isn't done. That one check is what makes the whole pattern a system instead of a habit: "am I the bottleneck right now?" is a query with a true answer, before the walk and after it.
The desk and the gate walk are the same idea wearing different clothes. Agents made drafting cheap, so the scarce resource is the moment a human judgment lands, and that moment repays purpose-built furniture. For documents, the furniture is anchored comments and immutable versions. For decisions, it's a queue you can trust, a package that respects your attention, and a record that makes each answer permanent. Either way the contract is the same: I say it once, it lands somewhere durable, and the work moves without me.