Key Takeaways

  • Watching an AI agent stream tokens line by line creates an illusion of control while wasting hours on micro-adjustments.
  • Frontier autonomous models left unbounded often invent irrelevant problems, such as chasing unneeded OIDC integrations or triggering false cybersecurity alarms.
  • Kerr's Law states that software team scaling is limited by keeping human mental models aligned with generated code, not by raw model intelligence.
  • Dave Kerr manages agents like Fable by treating them like junior engineers on eight-hour assignments rather than pair-programmers.
  • The 15-Minute Divergent-Convergent Agent Workflow gives developers a disciplined way to set boundaries and step away from the terminal.

The 15-Minute Divergent-Convergent Agent Workflow

Dave Kerr built this method after finding himself trapped in endless prompt micromanagement. As he explained, he had to forcibly break out of the habit to “tweak, tweak, tweak, tweak, tweak, instead kind of point the bazooka, let it go.” (play from 1:44:49)

Step 1: Morning Divergent-Convergent Planning (15-20 Minutes)

Conduct a focused morning conversation with the agent. Start divergent by exploring the problem scope: what are we doing? Then rapidly converge into an agreed-upon, actionable execution plan that outlines the exact steps needed to execute the project.

Step 2: Unattended Execution Block (8 Hours)

Hand over the execution plan and leave the agent to work independently for the entire day without intervening. Explicitly instruct the agent to generate tangible interim artifacts, resolve roadblocks autonomously, and build interactive touchpoints. Give the system explicit parameters: “Okay, you basically have all day. I'm not coming back for eight hours. I need you to create interim results, be creative, work through problems. I need something I can touch, and interact with, and feel.” (play from 1:44:23)

Step 3: Evening Review and Overnight Iteration (15 Minutes)

Reconvene at the end of the day for a brief structured review of the artifacts. Check the results against your mental model: review what worked, what failed, and what alternative paths exist. Provide high-level guidance for the agent to continue iterating overnight before starting the cycle again the next morning.

When This Works (and When It Doesn't)

This workflow succeeds when building software features with clear architectural boundaries and verifiable end states. Kerr designed it to break developers out of unproductive prompt micromanagement during multi-hour, long-horizon autonomous software projects. It shines when you need interactive prototypes, CLI tools, or background service scaffolding where you can evaluate the finished artifact by touching and running it.

It breaks down when an agent tackles open-ended architectural research or ambiguous system integration. As Adam Stacoviak pointed out during the conversation, unbounded frontier models like Codex GPT-5.6 Sol Ultra or Fable will “just sort of go until they'll almost even invent problems to keep going.” (play from 1:49:32) Without rigid guardrails, an agent left alone for eight hours will happily spend four of those hours wiring up complex OIDC authentication you never asked for or spinning its wheels against aggressive cybersecurity false alarms. If the task requires deep domain context that cannot be validated with concrete behavioral tests, eight hours of unattended generation will leave you with a codebase that violates Kerr's Law: you will spend more time rebuilding your mental model of what the agent wrote than it would have taken to write the code yourself.

What to Do With This

Pick one self-contained feature on your backlog this week, such as an internal CLI command or an isolated webhook handler. Spend tomorrow morning between 8:45 AM and 9:00 AM running the divergent-convergent planning phase with your agent. Spend the first seven minutes exploring edge cases and error handling, then spend the remaining eight minutes forcing the agent to output a numbered, sequential build plan with defined interim test artifacts. Close the terminal window at 9:00 AM, set a calendar alarm for 5:00 PM, and do not reopen the agent session until the alarm rings.