How I run a team of AI Agents 24/7

Ryan Carson, a solo founder with roughly 25 years of CEO experience including scaling Treehouse to 110 employees and one million coding students, now runs an AI divorce agent called Untangle for family law firms with one employee. He projects revenue to 4x in a single month and argues that people who know how to run teams of AI agents will outperform everyone else in the current AI age.

The central operational claim is that cloud-based agent environments are essential to shipping at high velocity. Local development cannot scale to running five, ten, or more parallel agents because code copies collide and coordination overhead becomes unmanageable. Cloud agents such as Devin, Codex, Claude Code, AMP, Cursor, and Factory each spin up a fresh virtual machine per session, eliminating collision risk. Carson typically runs at least five cloud agents simultaneously, often ten, and ships an average of 22 to 25 pull requests per day, sometimes reaching 40 in a single day. On a day spent hiking with his son and largely off his phone, he still shipped approximately eight PRs before leaving.

Carson uses Devin as his primary agent tool, selecting it because its cloud environment is well-built and each session spins up a fresh virtual machine in the browser. He notes Devin launched roughly two and a half years ago but did not work well initially because the underlying models were not good enough, and that as of 2026 the models are sufficient for cloud-based harnesses to perform well. For orchestration he uses a Fable thread as a parent manager model and spins up child Devin sessions for individual tasks, substituting the cheaper Fusion model for child sessions because running a premium model like Fable continuously is not cost-effective. He does not give agents production write keys, instead storing them in 1Password and manually providing them only when needed.

Untangle runs several automated loops. End-to-end sign-up testing runs three times per week via a Devin playbook at approximately 60 dollars per run due to extensive browser testing. Devin records video of its own browser sessions, annotates them, and reviews the footage to identify and fix bugs it observes. A child triage Devin session triggers automatically when an automated test fails. A production watchdog runs every day at 9am, scans all database events for paid customers, and summarizes them in a JSON file linked to live UI so the founder can click through and see exactly what a customer was doing. Human pattern recognition from reviewing this output has caught UX problems that agent-built interfaces did not surface through logs alone. An agent named Grace acts as a paralegal chatting with attorneys and divorce clients, and a daily self-improvement loop grades her chats against a rubric, spawning a child session to fix any chat scoring below a threshold and open a pull request. Approximately three self-improvement fixes ship per day at roughly five dollars per session using SWE 1.7, a fine-tuned model built by Cognition specifically for coding, putting the daily cost of that loop at roughly fifteen dollars.

Carson spent approximately 20 thousand dollars in tokens last month, which he described as not viable, and estimates real engineering work per employee will settle at roughly five thousand dollars per month. He believes Claude Code and Codex are subsidizing token costs in a way that is not sustainable long term, and warns that building engineering workflows inside those tools risks lock-in to Anthropic or OpenAI models and pricing. Independent agent labs such as Devin, Factory, AMP, and Cursor are incentivized to optimize for affordable engineering by routing tasks across models. He recommends against building a custom in-house software factory unless at significant scale, and against relying solely on a frontier lab walled garden for a real product. A solo founder building a simple product can likely get by on a two-hundred-dollar-per-month frontier lab subscription, but once a real product and team exist an independent agent lab becomes necessary.

Carson describes the nature of work as having shifted to making 10 to 20 high-stakes decisions before lunch rather than two or three per day. He checks high-stakes agent threads approximately every 25 minutes to avoid mental exhaustion and does approximately 50 percent or more of his work from his iPhone using Whisper Flow for voice input. He acknowledges that even the smartest models still lack common-sense judgment about what is acceptable in a product experience, and that there is currently no substitute for a human founder personally using their own app to catch UX and obvious quality issues. He argues that using agents makes non-technical founders more technical over time through hands-on exposure to concepts like Postgres, migrations, and development versus production environments. He cautions that the tools and practices he describes will change every three to six months. His forward-looking view is that agents will eventually write, review, and ship 100 percent of code, and that everyone will end up operating some form of software factory where agents handle all coding tasks.