Zack Proser's four-layer stack keeps agents shipping while you step away
WorkOS engineer Zack Proser describes a four-layer system — signal agents, verification gates, self-improvement loops, and biometric awareness — designed to let AI coding agents keep working while the developer steps away, addressing burnout from constant context-switching.
Score breakdown
The framework reframes the AI coding bottleneck from tool speed to developer attention, and proposes concrete automation layers that allow agents to run and self-verify without requiring the developer to remain at their desk.
- 01Proser works on the Applied AI team at WorkOS, which provides drop-in APIs for enterprise go-to-market.
- 02He identifies human attention — not agent capability — as the hard constraint in AI-assisted development.
- 03His workflow uses voice briefings at 184 words per minute to dispatch agents to isolated git worktrees.
Zack Proser, on the Applied AI team at WorkOS, opens by describing a familiar pattern: developers using AI coding agents are getting more done than ever but are exhausted by 11am, burned out by constant context-switching and adrenaline. His concrete example is a bug fix for a Slack bot he built at WorkOS — a tool that lets anyone in a Slack channel generate a properly formatted blog post. When a colleague reported that the sentence-case enforcer was mangling acronyms like "SSO," Proser gave Claude Code MCP access to both Slack and his Linear tickets, instructed it to fix the bug and verify its own work without stopping, and let it run. Claude fired a test post into the blog channel, the bot processed it end-to-end, and Claude confirmed the fix — returning a completed loop with no further human intervention required.
From that experience, Proser built a broader framework for sustainable agentic development.
From that experience, Proser built a broader framework for sustainable agentic development. The four layers are: (1) signal agents that continuously read Slack and Linear so the developer avoids opening those tools directly; (2) verification gates that escalate from lint and build checks up to browser click-throughs and critic passes; (3) a weekly agent pass over JSONL conversation history to identify inefficiencies and generate missing skills; and (4) biometric integration via an Oura ring connected through MCP, allowing Claude to surface sleep data before a session begins. The operational pattern pairs voice briefings — delivered at 184 words per minute — with agents running in isolated git worktrees, while the developer monitors progress remotely from a phone on LTE.
Key facts
- 01Proser works on the Applied AI team at WorkOS, which provides drop-in APIs for enterprise go-to-market.
- 02He identifies human attention — not agent capability — as the hard constraint in AI-assisted development.
- 03His workflow uses voice briefings at 184 words per minute to dispatch agents to isolated git worktrees.
- 04Signal agents read Slack and Linear on a loop so the developer never has to open those tools directly.
- 05Verification gates escalate from lint and build checks up through browser click-throughs and critic passes.
- 06A weekly agent run over JSONL conversation history surfaces inefficiencies and generates missing skills.
- 07An Oura ring is connected via MCP so Claude can surface sleep data before a development session.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 12, 2026 · 10:05 UTC. How this works →