How I manage 15 AI agents 24/7 as a solo founder | Ryan Carson

Ryan Carson describes running roughly 10–15 concurrent Devin threads while building Untangle, software for family law firms. His constraint is managing priorities and reviewing work, not generating more code. He and the host also argue that customer conversations remain necessary to decide what is worth building.

Evidence basis: Summary of the saved English auto-generated YouTube captions, with obvious product-name spellings normalized here; no original-audio check was performed. Agent counts, spending, throughput, and product outcomes are speaker-reported, not independently verified. Carson discloses receiving $20,000 in monthly Devin credits after approaching Cognition; his preference for cloud agents is an opinion, not a comparative evaluation. 09:16–10:20(opens in a new tab)

Prioritize work and surface problems

• Use folders for work and a paper list for attention. Carson groups threads into bugs and P0/P1/P2 priorities. P0 holds the business initiatives that must advance despite interruptions; a separate paper list of weekly priorities keeps those goals visible while multiple screens and agents compete for attention. 10:20–14:03(opens in a new tab)
• Make account monitoring answer actionable questions. His Watchdog playbook repeats an account check for each customer firm: activity since the previous run, errors, and Sentry reports. It surfaces the top three problems and checks whether each is already fixed, being worked on, or associated with an unmerged PR. Carson then investigates through the same thread. He describes running it when he needs an overview, not a demonstrated continuous monitoring service. 16:05–19:02(opens in a new tab)
• Delegate business operations as well as code. Carson uses an investor-update skill; the host describes Devin handling custom quotes, documentation, and customer triage. The host's framing is a colleague who knows the codebase and can write code to solve business problems, rather than an agent restricted to implementing features. 13:12–15:06(opens in a new tab), 22:42–23:45(opens in a new tab)

Keep verification and merge authority explicit

• Carson's “land PR” retains a human approval step. Once an agent reports a PR ready, he triggers a playbook that starts a fresh Devin Review and allows up to two review-and-fix loops. The agent then records a browser walkthrough with captions and test results. It must show Carson that video; he says “video approved, land it” before it merges. He also reports using GitHub merge queues. This is not a claim that review loops alone make unattended merges safe. 27:56–30:43(opens in a new tab)
• The host's automated approval is a different workflow. Her “Merge Mommy” runs after CI checks, including Bugbot, pass. It scores PR risk, approves low-risk changes so they can merge without human approval, and sends medium- or high-risk changes to a human through Slack. The discussion does not establish that those risk scores are accurate. The host also reports inconsistent success with Devin video verification, contrasting with Carson's experience. 26:59–29:01(opens in a new tab), 29:56–31:34(opens in a new tab)
• Local pairing still has a role. Carson prefers Codex for UI-heavy work that benefits from low latency, a local browser, and occasional direction. The host describes a concrete verification loop: write user stories, open the preview branch in Chrome, exercise it as a user, report which stories pass, and fix discovered bugs. These are reported practices, not evidence that browser checks replace other testing. 23:45–26:59(opens in a new tab)

Preserve context without letting output choose the product

• Keep team discussion near agent work. Carson proposes collaborating in Devin threads so context is not scattered across agent sessions, Slack, and documents. The host favors public Slack threads because colleagues can observe real coding work and learn what agents can do. Her argument is about adoption and visibility; his is about keeping task context together. 30:43–35:51(opens in a new tab)
• Carry design decisions into implementation. Carson values Claude Design for extending an existing design system. The host describes creating a design file there, downloading it, and giving it to Codex alongside Figma material to build shared components in a monorepo. She also tested pulling the resulting tokens back into Figma components. This is the host's handoff workflow, not a demonstrated guarantee of design fidelity. 35:51–37:45(opens in a new tab)
• Customer learning still determines what to build. Carson says Untangle pivoted from consumers to family law firms after a lawyer identified a discovery-work problem and willingness to pay. After landing the customer, he asked to visit their office. Both speakers reject equating more generated code with more valuable product; the host reserves overnight work mainly for bugs and triage rather than asking agents to invent and ship features continually. 05:37–07:20(opens in a new tab), 18:01–21:43(opens in a new tab)
• Hire for observable agent-management work. Despite the solo-founder framing, Carson plans to hire. His initial screen asks candidates to record their whole desktop while adding a feature to an existing app; shortlisted candidates would then build something with access to Devin. He presents this as an experiment whose hiring results are still unknown. 39:38–41:32(opens in a new tab)