Ryan Carson is the solo founder of Untangle, a B2B SaaS platform for family law firms. In one month, he spent $20,000 on Devin.
For that money, he runs 10 to 15 agent threads at once, sorted into four folders: P0, P1, P2, Bugs. He ships up to 40 pull requests a day. He works across eight screens. He uses Devin for engineering, QA, customer triage and investor updates. When a feature is visually complex enough that he wants to stay close to the implementation, he uses Codex.
One founder, running a development team's worth of throughput on a credit card.
He says that setup did not lead him to product-market fit. One meeting with a family law attorney named Renee Bauer did. He listened.
Put those facts next to each other. The useful thing is in the gap. Agents collapsed the cost of building to almost nothing. The cost of knowing what to build is still one uncomfortable conversation at a time.
That asymmetry got worse. When execution was slow, being wrong became visible early. Now you can be wrong at forty pull requests a day, on schedule, with excellent test coverage.
The useful question about Carson's setup is what stops a dozen-plus tireless agents from shipping the wrong thing very efficiently. His answer is three small numbers, all chosen by a human in advance.
The number that gets quoted, and the number that matters
Forty pull requests a day is the number that travels. It is a good number. On its own, it is meaningless. A pull request is a unit of activity, not a unit of progress, and an agent fleet is extremely good at manufacturing activity.
Carson knows this, which is the interesting part. Both he and Claire Vo, who runs the How I AI show where he laid this out, are explicitly sceptical of letting agents run unconstrained overnight. Cost is not their stated reason. Their reason is that frontier models will generate an enormous amount of output and have no idea what customers actually need.
Here is a founder who built an unusually large agent operation for a company of one. His own account of finding PMF contains no agents at all.
The two accounts fit together. They show a division of labour, and it is worth being precise about where the line falls.
What the fleet is actually for
Carson describes managing his threads the way a good manager treats direct reports. Each one gets a clear goal and the right priority. Then he avoids unnecessary hand-holding. The P0/P1/P2/Bugs structure is the thing that decides which of a dozen-plus simultaneous conversations gets his attention next.
Two of his workflows are worth stealing outright.
The first is a merge gate that does not involve reading the diff. When an agent says a pull request is ready, his "Land PR" playbook runs Devin Review on it. The loop runs up to twice, so the agent spends its own budget finding its own bugs first.
Then the agent records a video of itself testing the feature in a browser. It narrates the test cases with captions and shows a checklist of results. Carson watches the video. If it works, he types four words:
Video approved. Land it.
And the agent merges.
Think about what that solves. At 40 pull requests a day, no human is meaningfully reviewing 40 diffs. Anyone who claims otherwise is rubber-stamping. Moving the human check from the code to evidence of behaviour makes the volume survivable.
A narrated browser run is reviewable in the time it takes to watch. It also degrades far more gracefully than skimming code you did not write.
The second is a customer-success function that exists as one reusable skill. His "Watchdog" playbook walks every customer account and pulls activity since the last check, including Sentry errors, UX bugs and other anomalies.
It then filters the findings to the top three problems across the entire customer base. The result is explicitly not a raw data dump. For each problem, it checks recent development activity and reports whether a fix has shipped, is in progress, or has an open PR.
Whenever Carson feels the familiar anxiety of not knowing what is happening across his customers, he runs Watchdog.
The design decision is the constraint to three. An agent with access to your logs will happily return everything it found. Everything is not a report. It is a second inbox. Ranking down to three, each annotated with fix status, turns monitoring output into a decision.
The part nobody copies
The counterweight is deliberately low-tech.
Despite the eight screens and the dozen-plus threads, Carson keeps a handwritten list of the week's priorities on paper beside him. Agents constantly generate updates, questions and decisions. The physical list keeps him anchored to the three things that matter before Devin's overnight output pulls him elsewhere.
This is the Watchdog's top-three constraint, pointed back at his own week. In both cases, the agents produce more than a human can act on. The fix has the same shape.
That is the actual architecture. The fleet is the cheap part now. You can have a dozen threads running by tomorrow afternoon for a credit card.
What Carson has that does not come with the subscription is a pre-committed answer to "what am I doing today". No number of agents can renegotiate it.
Where the ceiling is
Which brings it back to Renee Bauer.
That conversation supplied the direction. Product ideas, priorities and market judgment, in Carson's framing, still have to come from a human who talks to users.
One disclosure changes how you should read the anecdote: Bauer is no longer an arm's-length customer. In July 2026, she joined Untangle as a strategic partner working on product design, brand and market growth. Her own firm now serves as a design and beta partner for Untangle's attorney-facing product.
The point still holds. The conversation came first, and the company moved because of it. But recruiting a design partner is the better description of a founder story where the customer ends up inside the company. It is not cold user research. Arguably, it is the more repeatable move anyway.
If you are a solo founder hoping the agents are the answer, put that framing next to the $20,000: direction comes from a human who talks to users. He was already running the fleet.
The fleet was not producing the insight. It produced throughput against a direction that had to come from somewhere else, and the somewhere else was a conversation.
This is the honest version of the story, and it is more useful than the triumphant one. The fleet was downstream of a decision it could not have made.
One more thing he changed
Carson's engineering interview is a screen recording. Candidates record themselves building a feature inside an existing application, whole screen visible, no introductory meeting. In the next stage, he gives them Devin access and reviews the replay of how they worked with it.
He finds this far more revealing than a behavioural interview, and the reason fits everything above: he is not hiring someone to write code. He is hiring someone to direct a fleet. The only way to see whether they can do that is to watch them do it.
What to take
Do not copy the forty. Copy the three.
Three priorities on paper before the agents wake up. Three problems in the Watchdog report, ranked, each with a fix status. Up to two review loops, then a video, then four words.
Every one of those is a small number a human chose in advance. They are the entire reason a dozen-plus concurrent agents produce a shipped product instead of a very fast mess.
Then go book the meeting. That part has not been automated. Carson, who has automated more of his company than almost anyone, is the one telling you so.
