AI Engineer has published Ryan Carson’s talk from AI Engineer NYC, given on 14 October. Carson, founder and CEO of Untangle, describes how the family-law startup builds software with cloud coding agents: one “engineering manager” agent, which he mostly speaks to by voice, coordinates up to about 20 builder agents through a ticket board.
Untangle sells an AI assistant to divorce and family-law firms for discovery follow-up, document collection and routine client questions. Carson said the product began as a consumer app and quickly moved to selling to firms. “We’re building an AI paralegal specifically for family law, focusing on divorce in particular,” he told the audience. On the company’s about page, he writes that he started Untangle after divorce touched his own family, including two of his sisters and his parents.
In the talk, humans still set priorities and speak with customers. The manager is a session in Devin that, Carson said, never writes code. It pulls work from a ticket board — Untangle uses Linear, though he said the specific tool doesn’t matter — and talks to child builder sessions. People don’t step into those builder sessions. The number of builders varies. He put the usual ceiling at about 20, and said it can run from about five to 25, limited by how well a person can follow what the manager is doing.
Because he doesn’t rely on the manager to check its children on its own, Untangle added an automation that forces it to read every builder session every 15 minutes. A slide in the talk showed a 10-minute interval; Carson said he’d settled on 15. “You have to build an automation. You can’t trust your manager agent.”
Each builder has a review agent, and Devin already includes reviews. Untangle caps that loop at about three rounds; any important issues still open after that go back to the manager, then to Carson if needed. A separate quality-assurance step spins up a child of the builder. Engineer Dale Spoonemore, who had recently joined, helped put that process in place. Carson said the merge queue is the point that clogs if too many agents are running at once.
Every builder also records a video of the feature being used, with tests passing. In Devin, Carson said, that’s a built-in test feature: it records the run and marks the tests green or red. The step he won’t skip is watching that video and deciding whether to ship. He said he was tempted, just before the talk, to approve a video from his phone, and went back to a laptop because he couldn’t see it clearly.
“If you’re coding locally, stop. You really need to be coding with a cloud agent.”
— Ryan Carson, in How One Founder Manages 20 AI Coding Agents at Once
He added that he didn’t mind whether teams used Devin or another cloud agent, and that staying on local development holds a team back because this kind of parallel session work isn’t possible there. Shawn Wang, who organizes AI Engineer, quoted the same two lines when the recording went up.
At a larger scale, Carson said each person typically has one engineering-manager session, sometimes two — he had two running at the time — and those sessions message one another. A separate “merge queue manager” session coordinates the flow of pull requests, including things such as database migration numbers, so agents don’t collide. Untangle uses GitHub merge queues. These manager roles, he said, are ordinary named sessions, not a separate product.
He also described a change in how Untangle hires. The company no longer starts with pre-screens or video chats, which he argued introduce bias. Candidates first send an unpaid, roughly 45-minute full-screen recording of themselves shipping a feature, usually a pull request, on an app they already maintain, with no instructions. A ranking agent watches the video and transcript and stack-ranks the applicants. Carson said he initially didn’t trust the agent’s first result, and later concluded it had been right.
The top five are invited to a paid 16-hour trial: they get a Devin seat, sign a non-disclosure agreement, and try to get a real pull request ready to merge. The work is recorded, in Untangle’s case with Loom, and reviewed by an agent. Carson said he also watched all 16 hours of those recordings in the last round, then spoke with the finalists only at the end.
He told the audience the workflow has been iterated hundreds of times and works well today, “and probably will be useless a week from now.” The recording is on the AI Engineer YouTube channel, and Untangle is at untangle.us.
Discover more from Sulat AI & Tech News
Subscribe to get the latest posts sent to your email.
