AI EngineerJune 11, 202625m

Your Attention Is the Bottleneck, Not Your Agents — Zack Proser, WorkOS

TL;DR

Your attention is now the bottleneck: Proser says agents can loop, verify, and scale far faster than humans can review, so the limiting factor in AI coding is judgment, focus, and recovery time, not raw agent capacity.
A tiny workflow change closed the loop on a real bug: By giving Claude Code access to Slack and Linear, he had it fix a sentence-case bug that mangled acronyms like SKIM and SSO, test the result through the Slack-based blog bot, and return only when the bug was actually resolved.
Voice-first coding compounds hard: Proser says he went from roughly 90 words per minute typing to about 184 words per minute speaking, which lets multiple agent sessions start running while a traditional prompt is still being typed.
Remote control changes what 'walking away' means: With Claude Code remote control enabled, he can leave a session running on his dev machine, message it from his phone over cellular, and review PRs from a trail or the woods without losing momentum.
Verification gates are the safety layer that makes speed usable: His stack ranges from lint, builds, and unit tests to browser-based self-checks and even agent-on-agent review inspired by Anthropic's constitutional AI framing.
The system should improve itself from your chat logs: Because Claude Code saves local JSONL histories, Proser runs or proposes scheduled analysis passes that look for repeated ambiguity, missing skills, or MCP gaps, then turns those findings into new skills, hooks, or tooling for next week.

The Breakdown

Four parallel coding agents can leave even Simon Willison wiped out by 11:00 a.m., and Zack Proser argues that the real constraint is no longer the models but your nervous system. His fix is not fewer agents, but tighter verification, fewer context switches, voice-first control, and workflows that let you keep shipping while walking the dog or standing in the woods with your phone.