Key Takeaways

  • Ryan Carson runs 10 to 15 concurrent Devin agents in the cloud to ship dozens of daily pull requests for Untangle, leaving local IDE coding assistants behind.
  • Engineering leadership has inverted: solo founders now spend their time delegating, organizing, and unblocking AI agents rather than typing code.
  • Claire Vo notes that mid-career management skills (like scoping deliverables and triaging work) matter more in modern software builds than raw syntax mastery.
  • Carson prevents attention fragmentation across dozens of active cloud threads by anchoring his digital queue to a physical Ugmonk pen-and-paper desk card.
  • This operational rhythm relies on Carson's Dual-Layer Agent Management System to separate immediate bug triage from core product roadmaps.

Carson's Dual-Layer Agent Management System

Carson runs his startup Untangle by treating autonomous AI agents like an engineering department. “If you're out there listening and you are doing engineering, you're doing your work locally, you really need to open your eyes,” Carson explains. “I think the future is pretty much 100% cloud agents.”

To prevent the chaos of 15 simultaneous cloud threads pinging him at once, Carson relies on a three-part operational architecture:

Layer 1: Cloud Agent Priority Bucketing

Structure all active agent threads inside the cloud environment into prioritized folders: P0 (critical business initiatives and major new features), P1 (important secondary tasks), P2 (lower priority backlog), and Bugs (immediate fixes and error retry investigations).

Layer 2: Analog Physical Anchor

Maintain a physical analog to-do card (using an Ugmonk pen-and-paper system) on the desk with top weekly priorities to avoid getting lost in multi-screen notification sprawl across agent runs.

Layer 3: Asynchronous Goal Assignment

Delegate high-level goals into specific thread folders rather than micro-managing step-by-step code edits, checking in asynchronously throughout the day via mobile or web.

As Carson puts it: “All of us have to uplevel our ability to manage agents. That is our job right now. How do you effectively manage 10, 20, 100, 500 employees basically?”

Vo agrees that managing AI is simply people management at machine speed: “The reason why I think folks of a certain level of experience are ready to cook right now with AI is all those management skills that you have developed over your career can be applied right here, right now.”

When This Works (and When It Doesn't)

This system shines when a solo founder or small engineering team scales past four or five concurrent AI tasks. At that volume, checking random browser tabs creates extreme context switching. Grouping tasks into P0, P1, P2, and Bug buckets lets you clear blockers by severity rather than whichever agent finishes first.

It breaks down if you lack a clear product spec. Autonomous agents burn compute and produce messy diffs when assigned ambiguous goals. If your problem requires customer discovery rather than clear software execution, throwing 15 cloud agents at the codebase just generates technical debt you have to review manually.

What to Do With This

Set up your cloud agent environment before starting work tomorrow. Create four folders: P0, P1, P2, and Bugs.

Write down your three core business milestones for the week on a 3x5 index card and place it under your monitor. When you open your agent dashboard, assign high-level feature tickets only to the P0 folder. When a bug report hits your inbox, dump it directly into the Bugs folder and trigger an agent run without watching the terminal. Check back every two hours, review the generated PRs against your physical card, and merge what works.