Try Pro free
Full episode 31:54

The best 6 min of 32

One Founder Now Runs a Game Studio With 20 AI Bots | Walden Yan

Walden Yan says Devin Bots can turn agents into proactive coworkers and shift engineers from reviewing code to setting goals and guardrails.

MMTS · Madison Kanna

The best minutes

  1. 17:14Agents can take ownership of measurable organizational goals3 min
  2. 24:33Code review can fade as teams build stronger guardrails3 min

Also worth watching

  1. 0:00One founder's 20 bots raised output to 1,900 PRs a week3 min
  2. 3:00Poke's proactive assistant design became engineering bots3 min
  3. 12:22Project context lets bots surface related work proactively2 min
  4. 20:26Agents may break organization-scale problems into experiments2 min

Skip about four minutes of sponsor reads, the custom-software discussion, and the ambition discussion and closing pitch.

Full analysis · 6 min read

Real Alpha · Nov 17

Walden Yan says weekly PRs rose from around 50 to 1,900 with 20 bots

  • Yan said a solo founder with 20 bots went from around 50 to 1,900 PRs weekly.
  • That is a striking throughput report, not evidence bots caused higher revenue or ratings.
  • Track accepted work, quality, and spend before treating one studio as a staffing plan.

Free with an account

Read the rest of the analysis

6 sections, the 6 other voices it tests Yan against, and what it means for you. Every episode we cut gets one.

Signed up before? The same buttons sign you in. The episode, its moments and the transcript stay free.

0:21

One founder's bot team multiplied reported PR throughput

Yan's example is a solo founder running 20 specialized bots at an award-winning game studio. She began at, he thinks, somewhere around 50 PRs a week. By the end, she was shipping 1,900 PRs weekly, he said.

Dividing 1,900 by the approximate 50 baseline gives roughly 38 times, our calculation. But the transcript gives no time span, so the increase's pace cannot be assessed. Nor does a PR count reveal how much each change does or how much rework remains.

For example, bots covered UI design and release management, while coordinating and holding daily stand-ups. That describes a broad operating setup, not simply more coding sessions running in parallel. The account still comes from one beta tester, with no other studio comparison.

Yan said the game's ratings and revenue increased over the same period. He gave neither the size of those gains nor evidence tying them to bots. Still, the case connects agent use with business outcomes, though it cannot establish causation.

Our read

The number is a useful signal of possible capacity, not proof of business impact. PR volume needs a quality and outcome measure beside it.

4:26

Cognition says bots already build Cognition's products

Yan said Cognition uses Devin to build Devin, and that its development process has moved to the agent. That is more relevant to engineering buyers than a launch demo, but it remains Cognition's own account.

He said merged PRs per bot message rose about fourfold after internal use of bots. That measures merged output per message, not cost per feature, defect rates, or time spent maintaining the result. So the ratio suggests leverage while leaving unit economics unanswered.

For example, one instruction set a bot to maintain Android parity with Cognition's iOS app. Yan said it ran and tested code, checked pixel comparisons, and merged thousands of PRs. The episode gives no acceptance rates or account of how often people corrected the work.

Yan co-founded Cognition, the company selling Devin, so these are vendor-reported internal results. Roman Ugarte said on Lightcone in November 2026 that he likes a bot dedicated to one problem. A separate Lightcone moment describes ad-account access and a budget, with a design bot helping test creative. Those examples support focused roles and bounded experiments, not a comparative performance result.

Our read

The internal examples make durable ownership plausible, but do not establish an external benchmark. Ask what shipped, remained reliable, and cost less.

17:48

Outcome goals can make agent spending governable

Yan wants teams to assign outcomes rather than individual tickets. His examples include faster application performance and higher onboarding retention, with bots organizing work around those measures. That shifts management from task assignment toward target setting.

He also says clear outcomes and budgets could make AI spending easier to control. But the episode offers this as a product rationale, not a reported deployment with verified savings. A budget without a stop rule still permits waste inside the limit.

A team could set a baseline, target, spending cap, and stop condition before delegating. For example, it could track onboarding retention while counting agent spend and human intervention. Then compare the outcome with a period before delegation.

Danila Shtan said on Unsupervised Learning in November 2026 that reliability matters for reaching outcomes. Alexander Wissner-Gross said on Moonshots that month that benchmarks can direct optimization toward chosen goals. Together, those points support measurement, not a promise that goal-based agents already work at scale.

Our read

Outcome-based delegation is plausible only when the outcome is independently measurable and spending is bounded. The useful test is whether the agent improves that measure, not whether it produces more work.

12:39

Shared context gives bots room to act proactively

Yan says proactive help depends on context built over time. He pointed to GitHub for project work and Slack for company plans. Without those connections, a bot may lack the signals needed to spot overlaps.

One example is a bot noticing another person working on a related project. Another is a bot flagging a change that affects the user's own work, then offering a fix to merge. Those are coordination benefits, not proof every suggestion is useful.

In November 2026, Lauren said on Lightcone that Slack connectors bring engineering agents into work context. Another moment on that episode describes a marketing bot with ad-account access and a budget. It paired that bot with a design bot to test creative.

That comparison shows why context and permissions belong in the same workflow design. But the MTS episode does not specify access limits, approval rules, or how irrelevant alerts are handled. Those omissions matter when bots can see team communications and act on code.

Our read

Context can turn an agent from a task runner into a coordinator, but it expands the control surface. Teams should treat access boundaries as part of workflow design.

24:49

Less review requires stronger system-level controls

Yan does not recommend ignoring AI-written code. He says less line-by-line review depends on better architecture, tests, and explicit guardrails. The engineering task shifts toward making classes of changes safe to automate.

His prompt example has bots test changes at least 10 times across various models. Changes beyond a certain length need human review, while some shorter changes do not. That is an escalation policy for prompt edits, not a blanket approval rule for all software.

Angela Jiang said on Latent Space in October 2026 that people still shape system design. Jake Stauch said on No Priors that month that background autonomy should grow through permissioned tools and workflows. Both comparisons keep human control in architecture or access, even as execution moves to agents.

The operating test is whether checks catch failures the team cares about. Teams should name which changes need approval, who owns exceptions, and how failures reach a person. Without those decisions, review reduction just hides risk rather than controlling it.

Our read

Review can move upward from individual lines to system design, but accountability does not disappear. Yan's own example depends on explicit tests and thresholds.

21:54

Yan expects agents to plan organization-scale experiments

Yan expects agents to move from engineer-sized tasks toward organization-level goals. He says bots could choose models for tasks and allocate resources across experiments. That is a forecast about orchestration, not a result the episode demonstrates.

Yan hopes that within a year, agents could break down a problem as broad as enabling new drug discovery. He said he doesn't think an AI can take on an entire organization-sized problem today.

Zach Lloyd described agents proposing changes to the factory that runs them on Latent Space in November 2026. That narrower engineering pattern uses benchmarks to compare system changes with real company tasks. It offers a more bounded step than delegating a whole organizational problem.

For a team, the practical sequence is to automate one bounded workflow, then measure quality, cost, and interventions. Widen its authority only when its controls hold across that workflow. The game studio's PR count is a reason to test, not a staffing forecast.

Our read

The episode points toward agents managing experiments, not just writing code. The credible path runs through measured results, scoped authority, and gradual expansion.

What it means for you

  1. Choose one outcome. Give an agent one measurable product or operations goal. Record the baseline and define what would count as improvement.
  2. Track accepted work. Measure merged and retained changes, not just PR volume. Compare quality and maintenance effort with the work the agent replaces.
  3. Set a spending limit. Pair each delegated goal with a budget and a stop rule. Review whether the agent reached the target within both.
  4. Write review thresholds. Specify which changes require human approval and which tests allow lower-risk changes to proceed automatically.

What to watch

  • Whether the game studio's higher PR volume corresponds to sustained quality and revenue gains.
  • Whether Cognition reports external results with cost, acceptance, and reliability measures.
  • Whether teams can safely widen agent authority while keeping budgets and escalation rules explicit.

Send this to the engineering leader setting agent budgets and review thresholds.

0:21

Watch the minute where Yan gives the weekly PR number

Written by Real Alpha from the episode's transcript and the other moments it links, then checked against the transcript. Quotes are the speaker's exact words.

Minute by minute

  1. 1:35Bots take on ongoing roles

    Yan says Devin Bots take on jobs, such as owning a release process, rather than single tasks. He describes a game studio run by one founder and 20 specialized bots: its output rose from somewhere around 50 PRs a week to 1,900, and Yan says the game’s ratings and revenue also increased.

  2. 3:00Poke inspired proactive agents

    Yan says Cognition acquired Poke, whose iMessage agent reminded users about things they had forgotten and helped them keep up with work. He says Cognition saw a chance to bring Poke’s personal, proactive and persistent agent technology to engineering teams.

  3. 4:05Cognition uses Devin to build Devin

    Yan says Cognition has long used Devin to build Devin and that he struggles to find code not written autonomously by it. For the Android app, the bot was told to maintain parity with the iOS app; Yan says it can go days without input while merging thousands of PRs and testing its work.

  4. 5:43Decisions reduce agent oversight

    Yan says the decisions feature helps people managing parallel work focus on issues that need their attention, while low-risk changes can be merged automatically. He says a bot can also start and manage a team of agents for a larger job, reducing the overhead of coordinating them.

  5. 10:09Agents can maintain custom tools

    Yan argues that people can build and maintain professional tools tailored to their work, rather than treating AI-built software as a toy. He says an agent can also handle scheduled, intelligent tasks itself, so a user may not need to build a separate app for every need.

  6. 12:22Proactivity depends on context

    Yan says proactive agents need to build context about a person over time, using sources such as GitHub and company Slack. He says bots have flagged colleagues working on related projects and prepared changes that address issues affecting a user’s work.

  7. 16:56Teams can set outcome-based goals

    Yan says engineers need to move from thinking like individual contributors to managing agents. Instead of assigning tickets, teams can give bots goals such as improving application speed or onboarding retention; he argues that tying outcomes to budgets can help enterprises manage AI spending.

  8. 20:36Agents may take on broader problems

    Yan says the next step is moving from individual engineering tasks to goals for an entire organization. He says bots can choose models for tasks based on price and performance, and hopes that within a year they can break down problems such as drug discovery into experiments, software needs and investment plans for people to evaluate.

  9. 22:35Guardrails could reduce code review

    Yan says code review may soon become unnecessary as engineers focus on architecture and guardrails instead of reading every line. He points to security-conscious code structure, formal verification and better testing; for example, prompt changes at Cognition must be tested at least 10 times, while some changes still need human review.

  10. 27:17Ambition can unlock more uses

    Yan says he was struck by a founder running a game studio with 20 bots and urges users to try more ambitious projects. He says Devin Bots fulfill his original goal of having an agent maintain 10 or 20 side projects, which he says is now possible with improved cloud infrastructure and multi-agent capabilities.

Published Nov 18, 2026.

Transcript · 27 min read

MTS, hosted by Madison Kanna, with Walden Yan.

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

  1. Walden Yan · 3 min

    “somewhere around 50 PRs a week at the start of it. She's shipping 1,900 PRs a week by the end of it.”

  2. Walden Yan · 3 min

    “Usually you would hire a team to go work on Android. Instead, we said, hey, look at our iOS application.”

  3. Yeah. And one feature that I liked when I've been using bots is the decisions feature. So basically it gets so overwhelming, like all of the different, you know, things you have to keep track of, like when is your agent done? Can you talk a little bit more about the decisions feature?

Read the whole transcript
  1. Yeah, no. Thanks for noting that. One of the things that people who are doing so many things in parallel notice is that there's a lot of context switching. There are a lot of things going on, and they really just want to know about what are the most important things. And so we have people where, yeah, you don't have to pay too much attention to the details of the conversation. You just know that, hey, if Devin bot actually needs me for something, it will give me an item to make a decision on. And so, yeah, I'm not going in, looking at the threads and reviewing every line of code. If Devin thinks something is important enough for me to review, I will talk about it high level at that decision point.

  2. And then for things that are low risk, I will just go merge automatically. And so it's the right level of interface for what actually needs me to look at versus what the agent can handle on its own. Yeah, as I was using it, I almost couldn't believe there was something, I couldn't believe there wasn't something like this before, even six months ago, because there's just so many times I'm lost in my threads. I don't know what to go back to, right? And just constantly forgetting when you're trying to use these agents. And so I really like that it's kind of telling me like, what do I need to look at right now versus I'm trying to keep track of everything in my head.

  3. Yeah, well, a year ago, if you look at people using Devon, And very few people would be trying to manage multiple threads in parallel. Come to today, many people are not only managing multiple threads in parallel, they're also trying to use devins to manage other devins. And so we realized that there's already this demand that didn't exist before for a better system to coordinate all these teams. And so this brings us to the other thing that really makes DevonBots different, I think, which is if you give it a goal or a job, the DevonBot itself will go spin up that whole team of agents and help manage it for you and basically reduce that overhead burden for yourself.

  4. Yeah, I think that's the part I'm still getting used to. Basically, I'm used to managing individual sessions, right? but then at this this abstraction level, like you have a bot managing other bots. I'm still getting used to it. What is your, what are your tips or heuristics for like working in this new way? I think that many people I've talked to when I first showed them bots, they are not ambitious enough because they're, they are so used to being bottlenecked by what the agent can do for them. I would really say, take your most ambitious ideas and you can keep throwing them at your devon bought because it can work on multiple of these threads and parallel and in fact it's it's better because over time it compounds and it learns uh from what things you want what things you don't want um when we were filming with venus williams actually uh there's actually a lot of complexity that goes into being a professional tennis player i learned that you know it matters what altitude you're playing at and humidity and you have to string your rackets to the right temperature or to the right uh pressure to match that uh the the team that they work with has to get the exact uh specification of balls that match the tournament balls ahead of time uh there's

  5. certain meal schedules and prep schedules you have to follow so there's actually a lot of things that go on behind her um and i think she wasn't used to before devon bots having something that could handle just every request that keep coming up but then i i really encourage her just call devon bots and just say what's on your mind. Say what else you need. And I think that was so powerful that you could just fully unleash your ambition. So yeah, my overall tip is don't hold back all the various things you want it to be able to do. You should just throw it all at it because it can handle it. I want to shout out our sponsor, Lovable.

  6. You know the app you've been meaning to build? The internal tool, side project, or product you'd otherwise lose a weekend to? With Lovable, you get software you can ship today. That includes an editable code base you can inspect and change, plus two-way GitHub sync. Lovable also handles the backend and infrastructure, including managed Postgres, auth, storage, hosting, payments, and more, so you don't have to wire it all up yourself. And through Lovable's MCP server, you can create and deploy your project directly from the agents you already use. Turn ideas into software people love with Lovable.

  7. Lovable.dev. Now, on to the episode. I'd love to talk more about this because this is actually something I noticed when I was first using my Devon bots um so I ran into this problem where you know I worked as an engineer for a decade and I think I'm used to doing things in an old school way like for example um I got into Devon I'm like okay build this app for me and then I realized I don't need to build an app like my bot Ada could just do this for me um and and then it was like from there I was thinking there are all these things where I'm thinking oh I need a whole app for this but really it's just my bot could do it entirely and run it, run it, run everything for me. Um, so I was thinking, yeah, if you could expand on that at all, like, how can I be even more ambitious with my bots? How can I be thinking more on like, how can I push the limits here more?

  8. The thing that I, let me actually walk this, this back a little bit. We filmed our launch video with a lot of people who are engineers, but a lot of people who weren't engineers. And I think some people initially may be surprised by that. We think that we're entering a world where it's not just non-engineers get to vibe code stuff. I think we're past vibe coding now. I think we're at a point where people can actually build and maintain their own professional tools custom to what they want to do. And it's not some toy thing. It's actually professional tools for the thing that they want to spend all of their time doing. When you get a Devon bot to work on it, you have this interesting mix where you get to consistently customize what you are doing, but the agent itself is also continually being intelligent and can help you fill in the gaps. So if you want another feature or idea or to do something, the agent can always fill that in for you and it can always keep on being proactive and staying on top.

  9. that for you. So I think your example was you were initially like, oh, I wish my app could message me on some schedule and be intelligent. It was like, actually, the agent can message you on the schedule and be intelligent. And it can still build apps and add that backing for you. But you never feel like you are limited by the software it's building. There's always that powerful thing behind it. Yeah, absolutely. And something else I liked about the bots is they

  10. Now that you've used Dev and Bots, I guess, for a little bit here, I'm curious, what are like kind of the biggest questions on your mind or maybe like some like, you know, what are the hardest hitting questions you have? We'll continue monitoring right after this message from our sponsors. ElevenLabs, AI that communicates at human level across every channel and modality. ElevenLabs.io slash MTS. Scale your startup on Neon. Millions of developers and startups have already chosen Neon for their backend. Start on the free plan or get up to $100,000 in credits for your startup at neon.com slash MTS.

  11. Special thanks to our sponsor Kong, the AI connectivity platform. Connect APIs, LLMs, agents, and systems with serious security and governance. konghq.com Finally prove what your AI investment is worth with Lairdyn. Head to lairdyn.com today to start measuring your AI's bottom line. MongoDB, trusted to run critical applications by 75% of the Fortune 100, the intelligent data platform for the era. Apply today for access to up to $15,000 in Atlas credits, dedicated technical support, and exclusive events and go-to-market opportunities. MongoDB.com. Chargebee, billing and monetization infrastructure for the AI era.

  12. chargebee.com ClickHouse is the real-time database built to keep up with every query your apps and agents send its way. And it now runs Postgres, observability, and AI tooling on one platform. Learn more at clickhouse.com You should fast. So do attackers. Don't assume you're secure. Prove it. Horizon 3. Go hack yourself with AI. visit horizon3.ai slash node zero the key ingredient in today's m&a srs aquium learn why their shareholder rep payments and escrow solutions have been used by 96 percent of top global pe firms and 88 percent of top global vcs at srs aquium.com slash mts that's s-r-s-a-c-q-u-i-o-m dot com slash mts i think it goes back to what i said earlier just like um i'm so used to all these individual sessions and uh one like the decisions future like getting used to that um but then also probably just like yeah a bot managing bots like i haven't done as much of that and so like shifting my mindset for like working in that way and and like figuring out

  13. Walden Yan · 3 min

    “You can actually give very clear and quantifiable metrics and goals to your bots and trust that they will go now spin up a team to go optimize that for you.”

  14. And it's funny to think about myself, you know, engineers switching from being software managers to thinking of yourself as a manager of agents. It just makes me think about the entire reason I worked as a software engineer was to avoid ever being a manager. But it's different because it's not humans. Anyways, so, yeah, I remember, I mean, Devin has been so early and you saw so many things.

  15. Walden Yan · 2 min

    “here's experiment one, two, three, here's the different pieces of software that we have to set up to make that successful.”

  16. I do want to quickly touch on, because I've had thoughts on this before about the changing role of an IC. I love to code. You know, I started this company in many ways because I love to code. I want to build a tool that makes coding easier.

  17. I think there was a period where we were in this uncanny valley of coding was getting a lot easier to write, or code was becoming a lot easier to write, but then you're doing the things you enjoy least, like reviewing code. um i feel us exiting that valley now i think actually devin bots is going to be a big part of this where i don't feel like i have to juggle so many threads and be reviewing the code that a different agent wrote for me i feel like i'm talking and working at this higher level now where i can actually just like let my imagination run free and i can still make the decisions i want the right architectures the right product decisions um and i get the satisfaction of clicking merge and seeing the PR merge and the confetti go off, I feel like I'm getting to a point where, wow, I'm reaching a new steady state flow that didn't exist for a little bit here in the age of AI. And that makes me hopeful that, you know, we're actually on this new ramp out where you can find this new thing to really love and enjoy about the way you code and build products today. Yeah, I completely agree.

  18. And it did feel when I was working as a software engineer that as I started using AI to code, it felt like a lot more of my job was just reviewing PRs, which again, not the most fun part. It's like coding and building something new and thinking about, thinking on these hard problems of how can I structure these things. And then it started to feel like I'm just, you know, it looks good to me on leaving that comment, like looking, sifting through someone else's code, the AI's code, which not fun at all. And yeah, I do feel like with the bots, It feels like this shift of now managing things, being super productive, less just, you know, reviewing everything in that kind of older way.

  19. Walden Yan · 3 min

    “And if you step back a little bit, the problem of how do I get my AI to not need me to review my code is actually a pretty fun engineering problem.”

  20. had that period as a software engineer where it's just started to feel a little depressing and now it's feeling more fun especially just the more you build and the more ambitious you try to be i think yeah the there's kind of a interesting thing of like the people who are now most ambitious i think will have a uh leg up with the tools of today and so i actually think jenny who i mentioned was running her own game studio with 20 dev and bots doing different roles and acting like you know actual uh roles of engineers at our company i was like wow i never thought of doing that or almost i was scared to push it to that level of ambition but she not only pushed it to that level of ambition she invested into making a real reality she made it happen with a

  21. real application making real revenue and users i was like wow that is that is uh fascinating and i hope more people will use this as an opportunity to, to try to do crazier things. Uh, I, I really do think that, um, it, it, it must be like a super, we, we must realize how super unique this time period is where we have these like new tools and toys that can do such powerful things, but you just have to point them in the right direction and, and, and give them the right ideas to start off what they're doing. Yeah, absolutely. And I remember some advice I was given when I was kind of having an existential crisis as a software engineer was like, just go build 10 crazy ideas. And then I think I quickly started going from scared of AI and scared of what was going to happen to getting really, really excited because it's kind of like, you know, the joke that software engineers have like, I have like 50 domains that I've never like had the time to build those projects, but they're all, I'm just hoarding those domains. And it's like, now is the time to use all those unused domains that we've all just been had in our, had for years. Yeah. I mean, this was, it feels like devon bots is realizing my original dream for devon which was i was just in my dorm room and i

  22. had like 10 20 different side projects and i wanted a ai that could just message and would go maintain those for me we ai was not nearly as good back then but we've come such a long way and i think we've built out the underlying cloud infrastructure and autonomous capabilities and multi-agent capabilities to actually make something like this possible so that's like super super satisfying um and yeah i think there's something unique too about having a devon bot that can actually go write code a lot of uh other tools out there are like personal assistants and sure they can connect to my email and run errands for me and do cool stuff i actually use a few of them um but one that i know has the capability of like a full software engineering force and i can just like throw arbitrarily large ideas at it and it won't you know fall behind And that's like, that's something that I'm really excited about with our DevonBots launch today.

  23. Yeah. And I feel like that is needed for absolutely everyone because it's kind of like you said earlier with the future of customized software and like just building any sort of app that works for you, like doesn't even have to be someone who is technical, right? I think that's great that it's good at that coding for pretty much anyone. Yeah, no, absolutely. Right on. Yeah. Okay, cool. Well, I think that's our time. Thank you so much for joining us. for someone who's just watching and they want to get started with dev and bots like what do they do first absolutely um you can just go to dev.ai slash bots uh sign up for for an account and you know play around with it awesome cool well walton thank you so much for taking the time to talk to me i've been following cognition for forever and i was really excited to talk to you so thanks so much for coming on mts thank you a huge thanks to mts sponsors blitzy autonomous software development for enterprise code bases. Ship 5X faster. Blitzy.com. AdQuick. Make your brand a billboard.

  24. Out-of-home advertising as easy to scale as digital. AdQuick.com. Arena. Measuring AI performance in the real world. Arena.ai. Support for the show comes from VCX, the public ticker for private tech. The U.S. stock market started history's greatest wave of wealth creation. From factory workers in Detroit to farmers in Omaha, anyone could own a piece of the great american companies but today our most innovative companies are staying private longer which means everyday americans are missing out until now introducing vcx a public ticker for private tech visit getvcx.com for more info that's getvcx.com carefully consider the investment material before investing including objectives risk charges and expenses this and other information can be found in the innovations fund prospectus at getvcx.com this is a paid sponsorship You

More interviews