Managing AI Agents: How One Founder Ships 25 PRs a Day, Solo

Ryan Carson runs a legal tech startup with real law firms as customers, a seed round in the bank, and revenue about to 4x this month.

Headcount: one. Him.

He ships 22 to 25 pull requests a day. Some days 40. Half of that work happens on his phone. The morning he climbed Mount Washington with his son, he shipped 8 PRs before the kid woke up.

The skill behind all of it is managing AI agents. Not prompting tricks, not "learning AI". Managing a team of agents the way an engineering manager runs a team of humans.

Ryan broke down his whole system on Greg Isenberg's podcast, and it's one of the most concrete walkthroughs of this I've seen. Here's the episode, then the playbook with the numbers.

Who is Ryan Carson (and why listen to him)

Ryan has been a founder CEO for 25 years. He's on company number four.

Company number three was Treehouse. It grew to around 110 full-time employees, taught a million people how to code, and got acquired. So he's done the classic version: hire people, build an org chart, manage humans.

His current startup, Untangle, is an AI agent for divorce law firms. Family law, the topic nobody wants to touch. Seed round raised, real paying customers, revenue on track to 4x in a single month.

And he's running the whole thing alone, by managing agents instead of employees.

His argument is simple: whatever you did before (people manager, engineer, VC, stay-at-home parent), your job now is manager of agents. The people who get world class at it will outrun everyone else.

Lesson 1: work in the cloud, not on your laptop

Ryan's first rule sounds harsh: "If you are working locally, I honestly think you are a caveman."

His reasoning:

When you code on your own machine, you can realistically do one thing at a time. Want to run two agents at once? You need git worktrees or a second copy of your codebase. Three or four agents? Forget it. You spend your energy untangling which change collides with which.

Cloud agents kill that whole problem. Tools like Devin (from Cognition) spin up a fresh virtual machine per session. Click a button, new environment, agent starts working. Nothing collides, and you never have to wonder if you pulled the latest code.

Ryan runs at least 5 cloud agents at once, often 10.

That's the unlock behind the 25 PRs a day. It's not one super-agent. It's 10 normal ones running in parallel while he reviews and directs.

He still does light front-end work locally (new UI needs human eyes on it fast). But the moment something can move to the cloud, it moves.

→ One person on a laptop is capped by the laptop. One person managing cloud agents scales horizontally.

I've seen this pattern on other episodes too. The forward deployed engineers making $1M a year are doing a version of the same thing: orchestrating AI systems instead of typing every line. I broke that down in this article about forward deployed engineers.

Lesson 2: your real job is 20 high-stakes decisions before lunch

Here's the part nobody warns you about.

When your output 50x's, your decision load explodes with it. Ryan puts it bluntly: you used to make maybe 2 or 3 high-stakes decisions a day. One important meeting, a couple of calls, then two weeks of execution.

Now? 10 to 20 high-stakes decisions by lunch. Every agent thread ends with a question only you can answer: ship it or redo it, approve the migration or stop it.

His system for staying sane:

• Pin the threads that matter. Out of 25 PRs, maybe 3 are the big ones for today. Pin those. Let the bug fixes rip in the background.
• Check on a 25-minute cadence. Clicking through threads all day fries your brain. Ryan checks his high-stakes threads roughly every 25 minutes and mentally rests in between.
• Keep a paper to-do list. He uses the Ugmonk analog system. Old school, but it keeps the 3 real priorities visible while 20 agent threads scream for attention.
• Work from your phone. Half of Ryan's work happens on his iPhone in a browser. In the shower, thinking "I should land that PR", then just doing it. His homework for Greg at the end of the episode: get to 50% of work done from your phone.

One more detail worth stealing: he never gives agents production keys. All prod credentials live in 1Password, and when an agent needs to write to prod, it has to ask him. He pastes the key in for that one session. Agents will eventually do something dumb, and that one habit caps the damage.

Lesson 3: build automations that improve your product while you sleep

This is where it stops being "coding faster" and becomes a software factory.

Ryan runs three standing automations at Untangle, and each one is a template you can copy.

1. The end-to-end signup test. Three times a week, an agent signs up for Untangle in a real browser, creates a case, onboards a fake client, and walks through the whole flow. It records a video of itself, annotates it, watches its own recording, and files fixes for the bugs it catches. Cost: about $60 in tokens per run. If signups break, he knows before customers tell him.

2. The production watchdog. Every day at 9am, an agent reads through everything his paid customers did in the last 24 hours and writes a summary with links to the actual sessions. Like a chief of staff briefing: here's what went well, here are two weird things, click here to see them. Ryan says this one sounds small and is actually huge, because you think you know what your customers are doing. You don't.

3. The self-improvement loop. Untangle has an AI paralegal named Grace who chats with attorneys and clients. Every day, an automation grades those conversations against a rubric. Anything below the bar kicks off a child agent session that fixes the problem and opens a PR. Ryan ships about 3 of these loop-generated fixes every day, at roughly $5 a session. His math: "If you're not willing to pay $15 a day to improve one of the core feature sets of your app, what are you doing?"

The pattern behind all three: pick something you'd normally check in a weekly meeting, and turn the meeting into an automation.

This thinking isn't just for code either. Follow-ups, lead routing, reporting, onboarding emails: the same logic applies to the boring ops in any business. If you want the non-software version of this playbook, this guide to business automations covers where to start.

I also wrote about the job-level version of this shift in how to automate your job with AI if you want to start smaller than "software factory".

The money part: what running an agent team actually costs

Ryan is unusually honest about the bill.

Last month he spent about $20,000 on tokens. His own verdict: too much, not viable. Where he thinks it settles: around $5,000 per month per "engineer equivalent" of real work.

How he's cutting the cost: model routing. The expensive frontier models handle the hard thinking. The cheap fine-tuned coding models (Cognition's SWE model, in his case) run the repetitive loops. That's how a self-improvement session comes out at $5 instead of $50.

He also uses a parent-and-children pattern: one smart model acts as the manager, spins up 5 cheaper child sessions, and coordinates their work. You don't run your best model on everything for the same reason you don't put your CTO on data entry.

His spicier take: the $200 a month all-you-can-eat plans from the big labs are subsidized, and they won't stay that cheap. Great for a solo builder tinkering. But once you have product market fit and a real product, he'd build on the independent agent platforms (Devin, Factory, Amp) because their whole business is squeezing the best output from the cheapest model mix.

You don't have to agree with that last part (I use the big-lab tools daily). The useful bit is the budget math: agents are cheap compared to salaries and expensive compared to a ChatGPT subscription. Plan for a real line item.

What I'd actually do with this

If you're a bootstrapped founder reading this, here's the ladder, easiest step first:

→ Step 1: run one cloud agent session this week. Codex, Claude Code on the web, Devin, Cursor's background agents, doesn't matter. Feel what it's like to delegate a full task instead of babysitting an autocomplete.

→ Step 2: run three at once. This is where the manager muscle starts. You'll feel the pull to watch them work. Resist it. Set a timer, check every 25 minutes.

→ Step 3: build one automation. Ryan's advice: literally tell your agent "I want to automate this task every X days" and let it design the machinery with you. Start with the production watchdog. It's the highest signal for the least effort.

→ Step 4: add a quality loop. Pick your product's most important surface, define what "good" looks like, and have an agent grade it daily and file fixes.

One non-technical note from the end of the episode that stuck with me: Ryan credits 20 years of sharing what he learns publicly (mostly on X) for most of his opportunities, including being on that podcast at all. Greg's old co-host Sahil Bloom started by writing about things he was learning, grew past a million followers, and turned it into a bestselling book and a fund.

Share what you're learning while you build. Ryan says it's the reason he was in that conversation at all.

FAQ

What does "managing AI agents" actually mean?

It means running multiple AI coding or work agents in parallel and acting as their engineering manager: assigning tasks, reviewing output, making the ship-or-fix decisions, and building automations so the agents improve your product on a schedule. The skill is delegation and judgment, not prompting tricks.

Do I need to be technical to manage AI agents?

You'll become technical by doing it, which is Ryan's actual point. He argues the "engineering is dead" take is backwards: good agent managers end up understanding databases, migrations, and prod versus dev environments because their agents surface those decisions daily. Start non-technical, stay curious, and the agents teach you the stack.

What are cloud agents and how are they different from Cursor or Copilot?

A cloud agent runs on a virtual machine in the cloud instead of your laptop. Each session gets its own isolated environment, so you can run 5 or 10 tasks in parallel with zero code collisions and manage them from a browser, including on your phone. Local tools cap you at roughly one task at a time.

How much does running AI agents cost?

Ryan spent about $20K on tokens in his heaviest month, and thinks sustainable spend lands near $5K per month per engineer-equivalent for a real product. Solo builders can start on $200 a month plans. The cost lever is model routing: cheap fine-tuned models for repetitive loops, frontier models only for the hard decisions.

Is this only for software companies?

No. The automation pattern (replace a recurring check-in with a scheduled agent that reports or fixes) works for lead follow-up, reporting, QA, and customer ops in any business. Software founders just feel it first because their whole product is text and code.