Descriptions:
Zack Proser, a developer on the Applied AI team at WorkOS, delivers a conference talk at the AI Engineer summit about a challenge that’s become common among developers using AI coding agents: the tools are extraordinarily capable, but the cognitive load of supervising them leads to burnout by mid-morning. His core argument is that developer attention — not agent capability — is now the true bottleneck in AI-assisted software development, and his talk proposes a structured workflow to address it.
The concrete example at the center of the talk: Proser used Claude Code with an MCP connection to Slack to fix a sentence-case bug in a Slack bot, then had the agent verify its own fix end-to-end — firing a test post into the channel, watching the bot pick it up, and confirming the output met the spec — all without human supervision. He describes a three-gate verification model ranging from basic lint, build, and unit test (gate one) to browser-click verification (gate two) to multi-agent constitutional review where a second agent checks whether the first met a defined constitution of requirements (gate three).
The proposed day structure is front-loaded: developers queue all tasks in the morning using deep focus, then launch agents across those work tracks and step away. Because Claude Code runs persistently on the developer’s machine and is accessible via phone, engineers can review pull requests and send new instructions while on a walk — reducing repetitive strain and mental fatigue while staying productive. Proser references Simon Willison’s observation about four parallel agents causing exhaustion by 11am as evidence this is a shared problem, not an individual one.
📺 Source: AI Engineer · Published June 11, 2026
🏷️ Format: Keynote Launch
