Ryan Carson, founder of family-law software company Untangle, describes his team’s cloud coding-agent workflow at AI Engineer. The agent counts and quality claims are his reports, not independently reproduced results.

One manager, many builders

  • Humans choose priorities and speak with customers. A manager-agent session reads a ticket board, chooses work by dependencies and assigns builder sessions; the manager does not write code.
  • Carson says a manager can coordinate roughly 20 builders, sometimes 5–25. His own ability to understand their progress is the limiting factor, not just the number of sessions the platform can start.
  • A scheduled automation tells the manager to inspect every builder about every 15 minutes. He prefers the explicit timer to trusting an agent to remember a check-in.
  • One ticket should produce one pull request. Builder, reviewer and QA agents iterate; significant unresolved findings escalate to the manager and then a human after roughly three review rounds.

The bottleneck moves to review and merging

  • A builder records a video of a feature in use and its tests. Carson reads the intended behavior, watches on a screen large enough to inspect it, and reserves the final ship decision for a person.
  • Adding more builders increases pull-request traffic. A separate merge-queue manager coordinates conflicts such as database migration numbering, alongside GitHub’s merge queue.
  • He advocates cloud agents and inter-agent messages. That’s a preference from his workflow, not proof that local development or simpler teams cannot work.

His hiring method is a separate, contested claim

  • Untangle asks applicants for an unpaid, roughly 45-minute recording of themselves shipping a feature in an app they already maintain; an agent ranks the submissions.
  • The top five get a paid, fixed-fee 16-hour trial on a real pull request under NDA. Carson says he also watches the recordings and calls finalists afterward.
  • Caveat: an AI ranking is not evidence of fairness or predictive validity. The unpaid screen can exclude people who cannot safely record existing work, and avoiding early conversation does not remove bias from the later model or human selection.

“You’ve got to watch the video … you’ve got to engage in as a product owner.” — Ryan Carson