Free with a sign-in.
The best 15 min of 43
Inside the Fastest-Growing Category in AI: Scott Wu, CEO of $26B Cognition
Scott Wu says independent AI coding agents can help companies turn model advances into software outcomes and unlock far more building capacity.
Sourcery · Molly O'Shea
- Scott Wu is cofounder and CEO of Cognition.
The best minutes
- 8:14Cognition negotiated the Windsurf acquisition in one weekend4 min
- 20:17Enterprises need help changing how they build softwarePro pick 7 min
- 33:24A MongoDB setup error revealed Devin’s potential4 min
Also worth watching
- 13:22Windsurf's product integration was gradual, not forced4 min
- 28:22AI progress still faces organizational and feedback constraints4 min
- 4:13Coding-agent adoption could reach far beyond software engineers4 min
- 37:13AI coding's bigger payoff may be expanding what people can build3 min
Skip the internal merger-planning discussion, the sponsor reads, and the closing event and newsletter plugs.
Full analysis · 6 min read
Real Alpha · Nov 8
Scott Wu says Devin writes roughly 95% of Cognition’s code
- Devin writes roughly 95% of Cognition’s code, while code shipped grew roughly sevenfold in six months.
- Those company-reported output measures do not show code quality or customer returns.
- Teams should track rework, delivery, and ownership alongside agent activity before expanding budgets.
Free with an account
Read the rest of the analysis
6 sections, the 5 other voices it tests Wu against, and what it means for you. Every episode we cut gets one.
Check your inbox
We sent a link to your address. Click it to finish signing up. Not there in a minute? Look in spam.
or
Signed up before? The same buttons sign you in. The episode, its moments and the transcript stay free.
Devin’s output rose, but output is not value
Scott Wu said Devin writes roughly 95% of Cognition’s code. He also said the company’s code shipped grew roughly sevenfold in six months.
For example, the first figure tracks code authorship, while the second tracks shipped volume. Neither shows whether work holds up after release.
Wu says Cognition measures productivity through delivered outcomes rather than tokens or spend. That is a better target than counting agent activity, but the episode provides no defect, rework, or customer-impact series.
These are company-reported internal measures, and Cognition sells the coding agent being measured. They are useful operating signals, not independent proof of net productivity.
On The Information’s TITV in October 2026, Meerah Rajavel, Palo Alto Networks’ CIO, reported threefold NPI growth over six months. She said no software engineers were added, and described a harness tracking quality, security, speed, and efficiency.
That comparison offers a measurement standard, not proof the companies’ results match. Rajavel links output to safeguards, while Cognition gives no equivalent series here.
“what you care about is outcomes”
Our read
Wu offers a concrete internal output signal, but not a complete productivity measure. Rajavel’s example shows how output can be paired with operating and quality measures.
Customer usage surged from an undisclosed base
Wu said customers’ Devin usage grew in the range of about 11 to 12 times. He placed that growth over the last six months or so.
He did not give the starting volume, customer count, or distribution across accounts. So the multiple signals activity, but cannot show whether adoption broadened or a few customers used Devin more.
Wu described coding-agent use a year earlier as roughly zero. He expects roughly every software engineer worldwide to use agents within two or three years, plus many people beyond engineering.
That is a forecast, not a market count. It assumes adoption, education, and organizational changes keep pace with product capability.
On MTS in October 2026, Aghi Marietti said enterprise AI adoption can jump after months of getting a workflow into production. That suggests adoption may lag, then rise sharply once rollout reaches production.
Our read
The 11 to 12 times figure signals momentum, but its missing baseline limits what buyers can infer. Wu’s forecast depends on usage spreading beyond current customers.
Enterprise adoption depends on workflow change
Wu said Cognition works across model providers, including OpenAI, Anthropic, and Google. He says enterprises need help changing how they plan, write specifications, design, and conduct user research.
That is a bigger offer than a coding tool. It also asks customers to change routines, assign ownership, and decide how agent work fits existing review and delivery processes.
Cognition sells Devin, so model neutrality is both product architecture and positioning. Wu says independent providers can focus on application-layer value, but the episode gives no customer comparison showing that independence raises returns.
Jake Stauch, on No Priors in October 2026, said Serval sends deployment strategists and software engineers onsite with customers. His example points to the human work involved in deploying enterprise automation.
For buyers, the relevant comparison is not just model capability. It is the cost and time of changing workflows against the value those workflows produce.
Our read
Wu’s strongest enterprise case is about implementation, not simply access to models. Buyers should count the change-management work rather than treating deployment as a software purchase alone.
Integration moved gradually after a weekend deal
Wu said Cognition and Windsurf worked out their acquisition over one weekend. Discussions began Friday evening, with an announcement ready by Monday morning.
The companies brought different strengths. Windsurf had an IDE, enterprise sales team, and customer relationships, while Cognition had its coding agent.
The scale mismatch was substantial. Cognition had 35 or 40 people, while Windsurf had around 200.
Wu said the teams first focused on serving customers and keeping both products moving. They integrated features gradually, letting overlapping use and new capabilities pull the products together rather than imposing a deadline.
That approach matters because strategic fit does not guarantee a smooth operating transition. Wu described billing, customer relationships, product handoffs, offices, and team coordination as details to resolve.
He said Devin Cloud and Devin Desktop now work together, with sessions moving between them. But the episode gives no retention, migration, or customer-satisfaction figures to test the integration’s success.
Our read
The deal shows fast strategic execution followed by slower product integration. Its fit is plausible, but Wu’s account does not quantify customer results.
One debugging story shows possibility, not reliability
Take the MongoDB setup story. At the end of 2023, cofounder Walden could not get the database running, so he asked an early Devin prototype to fix it.
Wu said the agent used broad technical knowledge and command-running ability to diagnose the problem, then got the setup working. It ran commands for a few minutes before reporting success.
The episode presents this as the moment Cognition believed an AI engineer could do useful work. But one memorable setup failure is a demonstration, not a reliability rate.
Jad Naous, on a16z Podcast in October 2026, said a data-lake feature took maybe about ten days. He compared it with maybe about six months for one person, while human review tied parallel work together.
Danila Shtan, on Unsupervised Learning in November 2026, said employees remain responsible for agent-generated code. Nebius keeps fully autonomous agents away from core services for now.
Together, these examples put productivity beside verification. A working task or faster feature does not settle how much review a team needs in production.
“the encyclopedic knowledge and the ability to actually go and run commands”
Jad Naous · a16z Podcast · Oct 16AI cut a complex data-lake project from six months to ten days- DSDanila Shtan · Unsupervised Learning · Nov 2Nebius lets engineers use AI but keeps accountability human
Our read
The anecdote explains the product’s appeal, but not its operating risk. Other teams’ examples keep human review and accountability in the measurement picture.
Capacity, not speed, is Wu’s larger bet
Wu says the larger payoff may be building more, not doing existing work faster. He contrasts efficiency with capacity and asks what people could build if software became easier to produce.
For example, he names a DMV that works, smoother banking and medical-record experiences, one-off software, and self-driving software.
These are forecasts, not demonstrated outcomes. They suggest a case beyond engineering-hour savings, but Wu gives no estimate for demand or product value.
Wu also doubts that a singularity arrives in six months. He points to organizational processes, hardware limits, and feedback delays.
He said he thinks AI will eventually solve intelligence problems, while acknowledging that amorphous work can take years to evaluate. A simple program can be tested quickly, unlike a ten-year strategic decision.
That distinction matters for budgeting. Software with clear tests may show gains sooner. Long-horizon decisions can remain hard to judge.
Our read
Wu’s abundance thesis asks companies to test what new software becomes feasible, not just labor saved. The episode supplies no demand estimate, so those possibilities remain a bet.
What it means for you
- Measure production quality. Compare agent-written work with existing defect, review, and rework measures. Keep shipped volume separate from production results.
- Ask for the baseline. When reviewing usage growth, ask for starting volume, active users, and spread across teams. Those measures show whether usage is broadening or concentrating.
- Budget rollout effort. Estimate the time needed to change planning, specifications, and review workflows. Compare that effort with delivery outcomes, not token volume alone.
- Test capacity gains. Choose one product or internal workflow that is currently difficult to build. Measure whether an agent makes it viable and whether people use it.
What to watch
- Whether Cognition reports quality, rework, or customer outcomes alongside shipped code.
- Whether Devin usage growth persists beyond the reported six-month window.
- Whether gradual product integration produces customer retention and broader adoption.
Send this to the engineering leader setting AI productivity measures.
Watch the minute where Wu gives Devin’s internal code share.
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
A Pro write-up, open to you here. Try Pro free for 7 days
A Pro write-up. The minute by minute is in Real Alpha Pro; the episode, its moments and the transcript stay free.
Read it with Pro, 7 days free- 1:55Outcomes matter more than tokens
Wu says Devin writes roughly 95% of Cognition's code, while the company's code shipped grew roughly 7x in six months. He says tokens are a poor productivity measure; Cognition tracks shipped outcomes, product progress and business KPIs instead.
- 4:13Adoption extends beyond engineers
Wu says customers' Devin usage rose 11–12x in six months. He sees a market that could grow from roughly zero coding-agent use a year earlier to every software engineer and many others in two or three years; O'Shea notes Cognition reached $500 million in revenue in under three years.
- 8:14Windsurf complemented Cognition's agent
Wu says Cognition saw an opportunity after learning some Windsurf researchers were going to Google. Windsurf brought an IDE, customers, enterprise sales expertise and platform engineering that complemented Cognition's agent; Wu says they negotiated the acquisition over a weekend and had the announcement ready Monday.
- 13:22Team integration came before product consolidation
Wu says Cognition had 35 or 40 employees and a $70 million or $80 million revenue run rate, while Windsurf had around 200 people. They first reassured customers and kept shipping, then used offsites and strategy sessions to bring the teams together and moved into a San Francisco office that fit everyone.
- 15:07Products merged through gradual use
Wu says Cognition kept both products going instead of forcing an immediate merger. Over time, Devin Cloud and Devin Desktop became more connected, letting developers start a session in one and hand it off to the other as users sought both local IDE and cloud-agent workflows.
- 17:24Founders handled the deal hands-on
Wu says the acquisition had no formal playbook: Cognition's founders and functional leaders worked with Windsurf's team on product commitments, integration steps and go-to-market. He says around 30 of Cognition's first 56 employees had founded a company before.
- 19:57AI should remain broadly accessible
Wu says Cognition values independence and works with multiple model providers. He sees government as a major opportunity for software, pointing to the DMV as an example of a service held back by archaic systems; he also wants AI policy to avoid concentrating control in one entity.
- 25:19Enterprises need more than model access
Wu says enterprises want long-term partners because it's hard to know which model will be best in 6 or 12 months. They also need help changing planning, specifications, design, user research and workflows; simply handing teams a chatbot is not enough.
- 28:22Wu sees six months as too soon
Wu says six months is too short for a singularity. He points to organizational and hardware constraints, as well as tasks with unclear success criteria or long time horizons, while arguing that AI will eventually solve intelligence problems.
- 32:31Wu says AI hype needs perspective
O'Shea says she found claims in a Colossus piece hard to accept. Wu says AI could reshape lives more fundamentally than mobile phones or the internet, while urging people to stay sane.
- 33:24A MongoDB setup proved Devin's premise
Wu says the team was experimenting with coding agents in late 2023 when cofounder Walden struggled to set up MongoDB. Devin could draw on knowledge of possible errors and run commands to diagnose them, then fixed the setup; Wu calls it the first moment the team believed an AI engineer could do work for someone.
- 37:13AI could expand capacity, not just efficiency
Wu says the bigger opportunity is not only doing existing work faster, but building much more. He imagines software generated on the fly, including single-use software, and asks what people will create when software can drive itself.
- 39:50Poker was proposed as a deal tiebreaker
Wu says he proposed a heads-up poker match with Peter Thiel to settle a gap in the terms of an early Cognition round. Thiel declined, and they worked out a deal through negotiation instead.
- 41:17Abundance could free time for priorities
Wu says he looks forward to AI handling personal decisions and routine details so people can focus on what matters to them. He expects the abundance era to arrive soon.
Published Jul 27, 2026.
Transcript · 38 min read
Sourcery, hosted by Molly O'Shea, with Scott Wu.
Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.
Cognition has raised over a billion dollars. First AI software engineer, Devin. Microsoft and Cognition. Cognition and Mercedes-Benz. Goldman Sachs, Mercedes-Benz, Citi, Dell, Santander, NASA, the U.S. Navy, and the U.S. Army. They've raised over 2.5 billion today at a last valuation of 26 billion. In less than three years, you got to 500 million in revenue. That's crazy. Half a billion in revenue in less than three years. Our customers in the last six months or so have also grown their usage in the range of like 11 to 12x. Once you've solved AGI, you just take over the entire world, and then there's just one entity that pulls the world.
I just don't think that's the right future for it. There is nothing that humans do that AIs can't do. Maybe you can just have like an AI engineer buddy that can just do work for you. But that was like the very first moment that we believed it was possible. I think the Abundance Era is real, and I think we'll do that for sure. Devin welcome to sorcery yeah do you get that a lot um yeah yeah yeah like I'll be in the elevator and in my like building and like I'll just come down back are you Devin and then you know like well I'm actually I'm actually not well Scott Welcome to Sorcery. We are in Paris, which is pretty cool. We're here for the Rays Summit.
It's like this huge AI Summit. I think this is like their second or their third year. You're going to be on stage? Yeah, I'll be on stage tomorrow. I just landed an hour ago and came straight here. Well, it's awesome to have you on. I'm excited to talk through the state of cognition and autonomous coding agents today. It's very pressing. Did you know that? I did, actually, yeah. I mean, I can't even keep up with it at this point. No, things have been going like crazy. I mean, even for us internally, one of the craziest things is, obviously, we use Devin to build Devin. In fact, we almost exclusively use Devin to build Devin at this point.
Read the whole transcript
Like, 95% of the code or something is written by Devin. But the crazy thing is, even in the last six months, if you ask people, oh, AI code, yeah, it's been going on for three, four years now, which is true, it has been. but even in the last six months i think the number for us was like our total amount of code shipped is roughly seven x in that time um and so it's like the number of prs we've shipped the amount of code um and obviously literal output you know like there's different ways how in terms of how you actually measure the productivity that you get out of it but i think in practice like we're just able to do so much more and get so much more leverage in even this this this period of time like the models just keep getting better i think there's more and more that's been kind of fleshed i would say with the product experience and so um yeah no things are things are flying it's crazy so how do you measure the productivity of it yeah it's a great question um so a couple things i'd say on this first of all i think obviously literal tokens is just not the right answer really i mean lots of people talk about that you know i think the token you know i kind of i say the token maximum era you know lasted from january to to may of 2026 roughly all right um yeah rip you know
It was a fun time. But no, I mean, I think we're now at this point where obviously what you care about is outcomes and what you're actually delivering. And so we see how much we're shipping. We see how much we're doing. Obviously, what do we care about is where are we getting to with the product? What are we actually delivering to our customers and our users? And I think that's roughly the same for every business. There's no kind of secret trick or there's no kind of easy way out in terms of measuring actual ROI productivity and so on, but a lot of it comes down to just actually looking to each of the outcomes that you're driving.
And so for us, it's, you know, obviously we have all. The KPIs of the business that we track and that we kind of follow. And, you know, when we think about our output, a lot of what we think about is obviously it's like, you know, how quickly are we growing those numbers as opposed to anything that we actually look at in terms of tokens or spend or something like that.
“the rough answer a year ago was roughly zero and the rough answer for you know two three years from now is roughly every software engineer in the world”
at half a billion in revenue and you have customers that entail i'm going to list them out because these are some pretty big customers so far goldman sachs mercedes East Benz, Citi, Dell, Santander, NASA, the U.S. Navy, and the U.S. Army. And you've raised over $2.5 billion to date at a last valuation of $26 billion. I want to go back in time because Cognition played a pretty big role
“we got in touch with the windsurf team that evening so it was friday evening and then basically over the course of that weekend”
This episode is brought to you by Brex, my favorite. You become what you spend on. And I refuse to spend my time on work that shouldn't exist. Expense reports, receipt chasing, and manual closes. The companies building what's next from Vercel, OpenAI, Anthropic, Granola, and Deepgram all made the same call. They all run on Brex.
brex is the intelligent finance platform that combines cards expenses and banking into a single stack with agentic finance built in ai agents that handle expenses automatically enforce policy before spend happens and close your books in minutes that's why sorcery runs on brex so i can spend time on building and not busy work it's time to get brex af learn more at brex.com slash sorcery that's b-r-e-x dot com slash s-o-u-r-c-e-r-y bye what was it like merging
“rather than force it and say, okay, over the next 30 days or 60 days, like we have to integrate these products.”
It's a good question. A lot of making it up. Okay. And I think for us, I mean, it's, one of the things about COG, honestly, is like we, and, you know, we've talked before about how, like, a lot of our team are former founders. Our first, like, you know, 56 people, I think like 30 of us were, had founded a company before this.
And so obviously, like, a lot of us together were people who liked being ambitious and entrepreneurial and thinking from first principles. But, yeah, also, I mean, I think the – no, I mean, it was a real, like, team effort. I specifically remember that weekend there was a point. So there was the initial kind of discussions, which were with, like, myself and Russell from our side and then Jeff and Graham, you know, from the windsurf side. But then, obviously, as soon as we got past, like, the very first initial discussion, there was a question of like okay but like what are we actually going to do and so then there was like every single kind of like part of the business that that saturday was it was one of my my favorite like honestly one of the funnest days of my life really um was like um and because you know we had our kind of like you know technical team we had steven and walden and so on who are like you know engineering product leaders like my co-founders going in like um you know sitting with their team to understand okay what's going on with windsurf what do we do to get out wave 11 what are the things what are the specific features which we need to make sure deliver on time, the things that we've promised
folks that we obviously want to live up to what are the next first steps on how we would start getting towards an integration, similarly on all of the go to market stuff, we had our folks there working with across teams, making that happen there's just kind of a feeling I think for all of us of like we just love figuring this kind of stuff out you know i think sometimes people kind of describe it as like oh like i just want to like build my product and piece and everything else about building a company is you know that's just like i'll put up with it but it's it's not my cup of tea i think for all of us like yeah like you know the experience of building the company is like is a lot of the fun of it and so it's a really fun weekend for us wow yeah i would imagine i'm sure it's something else every week so yeah yeah yeah it's fun i mean yeah we've had all sorts of uh yeah we've had we've had a lot of more fun weeks since that one too It's always a mess, yeah.
Well, speaking on the backdrop of M&A, with Cursor and their deal, their $60 billion deal with SpaceX, M&A is floating around and there's all those kinds of fun stories that are happening, but Cognition remains independent, so much so that you spent Independence Day at DC, in DC.
“I think for every company or for every team, like, what they care about is not how many tokens they're using, it's how much value they're driving”
If you're building what's next in AI, you need to know MongoDB, the database platform developers love and built for the agents you're running. MongoDB stores, searches, and reasons over your data in real time. with vector search and embeddings from Voyage AI all in the same system. No separate pipelines, no stitching together 10 different tools.
It's why 75% of the Fortune 100 and leading AI-native startups run on MongoDB. Build and scale from your first user to billions of vectors. Go to mongodb.com slash AI to learn more. That's mongodb.com slash AI to learn more. Bye. Assembly AI is a voice AI infrastructure layer millions of developers build on. They build the industry's best speech-to-text, voice agent, and speech understanding models that serve as critical... Infrastructure for companies like Granola, HeyGen, Ashby, and ClickUp. Their speech-to-text models lead the industry in accuracy and quality, and their speech understanding models help you go beyond transcription by uncovering insights, identifying speakers, and highlighting key information from voice data.
You can get started today at assemblyai.com slash sorcery and get $50 of free credits to start building voice AI products.
“i just don't think of that as like it's going to be all over in six months”
i think most of these people are already insane i don't know if you read the beginning part of your colossus piece yeah but what is up with the rest of the silicon valley tech community like i
read that and then i commented below and i'm like what the fuck like yeah i can't unread that thank you jeremy what was going on there there's some of that for sure no i mean it's um i it's it's it's crazy you know obviously it's like we're saying like these companies are growing faster before these capabilities are clearly like i think it's i think it's true that the ai paradigm is on another level compared to you know the mobile phone or the internet or something it's not just the technology that people use every day i think it will like fundamentally reshape our lives but yeah to your point it's just gotta stay sane so so the first task that devin did was
“but that was like the very first moment that we believed it was possible”
that's amazing thank you mongo db for yeah giving something engineers can go after yes thank you for the the complicated yet solvable documentation pages that was that It was just right in the cusp of what it needed to be in December of 2023. Oh, my gosh.
“the real unlock like, is not in efficiency, but in capacity, right?”
what are we going to create with all this abundance and i think i think we're like starting to get to that question so So... Last question, what was it like beating Peter Thiel in chess? Oh, I didn't beat him in chess, no. I would not recommend playing against Peter Thiel in chess. I don't think that's, like, a good, you know, you got to pick games that you can win. But I did challenge him to pull into a poker match.
I think there was, like, somehow there was, like, both, but it was just a poker thing that actually happened. There was a round that was one of the earliest rounds of Cognition, so this is, like, the seed, basically. And kind of got to a point, like, we were clear with that, you know, to their credit, by the way, they were always super, super upfront and super hesitant about how they negotiated with us, too. But we were at the point where we're like, look, we want to go and do this. Here's the numbers that would make sense for us. And they were like, here are the numbers that would make sense for them. And we were kind of close enough that you kind of felt like it should work in some form. It was just a question of, okay, what do we land on? And I was like, you know, since you're such a big fan of poker, we could just play a heads-up poker match for that and we could have the winner decide whose terms we actually signed it would have been for in retrospect it would have been for like one percent of the company or something which would have been a lot but but um um i put that with napoleon napoleon was kind of down and then peter shut it down oh peter you know i don't really think that's the right way to do this and then he said that you know and and so then that
didn't end up happening you know we worked it out anyway and we came to a deal and it was all good and then we've been been been working together since yeah what are you most looking forward to in the next 12 months yeah um you know i'm honestly just really excited about us getting closer to this era of abundance and i think we'll see a lot of that happen in the next like not just as somebody you know in ai but like as like a like as a spectator also like as a consumer like i i'm excited for ai to you know to be there for me for all of my like tough personal decisions that i have to go make i'm excited to have ai like handle all these different like little details in my life i'm excited to have like ai that like basically just like allows me to like spend all of my time and focus on the things that i care about and take care take care of the rest um i think like i think the abundance era is real and i think we'll be there pretty soon amazing such a positive optimistic way to end it thank you so much scott i really appreciate it thanks for having me and uh maybe we'll get some hot takes at Versailles.
Yeah. Okay. Thank you so much. Okay, cool. Huge thank you to the entire Raze team for an incredible event. And thank you to Brex, MongoDB, and Assembly AI for making this trip and series possible. If you enjoyed this conversation, you're going to love the rest of the Raze series with Tony Kim from BlackRock, Scott Wu from Cognition, Andrew Feldman from Cerebris, Rodrigo Yang from Salmonova, Michael Hurlston from Lumentum, cj desai from mongo db and many many more like our hot takes that we did at a secret location that you can find on x youtube and instagram subscribe to sorcery on youtube for more conversations with the people shaping ai and join the free newsletter you can also do paid at sorcery.vc for weekly insights on ai robotics enterprise software consumer semiconductors Did I say AI, AI again, and everything that's coming next, like funding announcements and all big things in tech.
Thank you. Bye.
