When agents scale infinitely, the bottleneck becomes your attention — a working developer's flow for not getting fried
Your Attention Is the Bottleneck, Not Your Agents — Zack Proser, WorkOS · Zack Proser
25 min total·Actually worth watching closely: ~7 min·2 must-watch clips
- 0:14 – 3:20Listen
Nuclear tools, ancient nervous systems
Opens with the thesis of the whole talk: agents scale infinitely via the cloud API and will loop until they meet the criteria you give them, so the bottleneck is no longer on the tool side. Human attention has a physical ceiling and only degrades under load.
Left unchecked, the default outcome isn't double the output — it's getting fried faster than ever before.
He's mostly just making the argument here, with a few title-style lines on screen; watching adds nothing, so this stretch is best in headphones on your commute.▶ Jump to 0:14Speaker · Zack Proser - 3:20 – 7:16Listen
Speed requires safety: three verification gates
The three layers that let agents run fast without wrecking things: hooks that verify lint, build and tests at the code level, telling the agent it must verify with the browser and actually click through, and a third closer to constitutional AI where a second agent checks the work against the constitution and hands back feedback to action.
The gates aren't constraints on the agent — they buy you back the time you'd otherwise spend watching it.
The three gates are laid out verbally one by one, with no flow chart or code demo on screen; catching the order and what each one covers is enough — write them into your own project config as you listen.▶ Jump to 3:20Speaker · Zack Proser - 7:16 – 11:30Listen
Voice-first: 184 words per minute, compounding
Introduces the three-layer stack for keeping balance — signal layers, voice-first, remote control — and focuses on voice: dictation regularly hits about 184 words per minute, which lets you prompt several agent windows in parallel.
Typing speed is a hard ceiling; once you dictate, the gap compounds over years of working.
He's describing his own habits rather than demoing, so there's no screen capture to watch — just listen; what matters is remembering how the three layers divide up, since everything after builds on them.▶ Jump to 7:16Speaker · Zack Proser - 11:30 – 14:17Listen
Fire off the work, then go for a walk
How remote control rewrites the working day: a deep focus session in the morning to queue everything into the agents, leaving the desk once they're on track, diffuse mode on the walk, and the insight going back into the session from your phone rather than waiting until you're back at the desk.
The walk isn't slacking off — it's the part of this process where the ideas come from.
It's entirely his spoken account of a day's rhythm, with no timetable or UI screenshot on screen; listen to it while you're walking yourself and it lands harder.▶ Jump to 11:30Speaker · Zack Proser - 14:17 – 17:29Listen
Making your harness smarter with use
How to turn one-off conversations into capability: review the moments where you and the agent had to go back and forth to eliminate ambiguity, ask what tools, MCP server or skills were missing, pull out the moments worth keeping, analyze them in aggregate later, and harden them into new skills.
The context of a work session is gold — every stall in the back-and-forth is pointing at the tool or skill you're still missing.
This is method and self-reflection, with no extraction script or dashboard on screen; trying it yourself once beats staring at the video.▶ Jump to 14:17Speaker · Zack Proser - 17:33 – 21:10Skim
Q&A: where skills come from
Audience questions about how skills accumulate through use, and how to mine worthwhile patterns out of past session data, answered from his own practice.
Skills don't get written once — they get forced out of you by the same pothole you keep hitting.
The camera mostly stays on the podium during Q&A and the questioners are off-mic, so the density is lower than the talk itself; run it at speed and slow down when a question interests you.▶ Jump to 17:33Speaker · Zack Proser - 21:10 – 24:37Listen
The night-shift agent, and a warning for newcomers
The end state he's converging on: Linear tickets for everything, subtasks for bugs and feature requests, tickets tagged 'Agent ready,' and a loop running every 15 minutes all day and all night, wired to Slack over MCP and the ticket system so nothing needs a human to check back. Then one hard reminder for early-career developers.
Don't use AI to do something you don't already know how to do — judgment and taste are the only thing you keep in this division of labor, and you have to build them by hand first.
This is where his position is sharpest; worth hearing at normal speed for the tone. It's still just him on screen, so there's nothing to watch.▶ Jump to 21:10Speaker · Zack Proser - 24:37 – 25:02Skim
Closing: three steps to start
Half a minute to land on three things you can do now: plug your highest-cost context switch — Slack, Linear — into your preferred pane of glass to build one signal layer, add the verification gates you don't have, then use the margin you get back to actually rest.
That last step, actually resting, isn't a pleasantry — it's the point of the first two.
A twenty-odd-second closing checklist; note the three items and move on, there's nothing on screen worth pausing for.▶ Jump to 24:37Speaker · Zack Proser