Devin | Sourcery Summary

1. Key Themes

From Zero to Ubiquity: The Coding Agent Adoption Curve

Scott Wu frames the market opportunity with striking simplicity: usage of coding agents went from essentially zero a year ago to a trajectory where every software engineer in the world — and many non-engineers — will be using them within two to three years. This isn't a gradual S-curve; it's a near-vertical ramp.

"The rough answer a year ago was roughly zero. And the rough answer for two, three years from now is roughly every software engineer in the world and probably a lot more than all just the software engineers because I think a lot more people will be able to produce code and produce products." 00:04:52

Dogfooding at Scale: Devin Builds Devin

Cognition's internal use of its own product is perhaps the most credible proof point in AI productivity. 95% of Cognition's own codebase is now written by Devin, and total code shipped has grown 7x in six months. This is not a marketing claim — it's the operating reality of the company building the tool.

"We almost exclusively used Devin to build Devin at this point. Like 95% of the code or something is written by Devin... our total amount of code shipped is roughly 7x in that time." 00:01:53

Customer Usage Growing 11-12x: Enterprise Adoption Is Accelerating, Not Plateauing

The most underappreciated data point in the episode is not Cognition's own growth but its customers' growth. Enterprises are not just onboarding — they are massively expanding usage after adoption.

"Our customers in the last six months or so have also grown their usage something in the range of like 11 to 12x." 00:00:25

The Token Maximization Era Is Dead: Outcomes Are the Only Metric

Scott Wu explicitly calls the end of the "tokens as productivity metric" paradigm, pinpointing its death between January and May 2026. What replaces it is pure outcome measurement — what are you actually shipping and delivering to customers?

"I kind of say the token maximum era lasted from January to May of 2026. R.I.P... what you care about is outcomes and like what you're actually delivering." 00:02:57

Independence as Strategic Moat: Model Neutrality as Competitive Advantage

Cognition has deliberately stayed independent and model-neutral, working with OpenAI, Anthropic, and Google simultaneously. Scott Wu frames this not just as risk management but as a philosophical bet that the value layer — not the intelligence layer — is where durable businesses are built.

"We're completely model neutral — we work with all the different providers. We work with OpenAI, Anthropic, Google, and so on... I think there is a lot of room, a lot of appetite from everyone out there for truly independent tools that they can use, and for things that really innovate at the product and the value layer rather than just the pure intelligence layer." 00:20:53

The Windsurf Acquisition: Speed as M&A Strategy

Cognition identified, negotiated, and announced the acquisition of Windsurf over a single weekend — Friday evening to Monday morning — with a team of just 35-40 people. The deal united an IDE (local) product with an async agent (remote) product, creating the full-stack coding workflow. The integration philosophy was gradual pull, not forced merge.

"We got in touch with the Windsurf team that evening. So it was Friday evening and then basically over the course of that weekend we worked out the entire acquisition together. So by Monday morning we had the announcement ready to go." 00:10:48

Government as the Single Largest Underserved Software Market

Scott Wu identifies government and defense — not enterprise SaaS — as potentially the single largest pool of unmet software demand on earth. The US Navy, Army, and NASA are already customers. The framing is not about contracts but about the fundamental mismatch between what government needs to build and the engineers available to build it.

"DC and government in general might perhaps be the single biggest kind of point of that — places that need software and that have things that they want to build but not enough software engineers to build them." 00:21:12

The Abundance Era: Efficiency Is the Wrong Frame for AI

Scott Wu argues that people are thinking about AI coding productivity backwards. The value is not in doing existing work faster — it's in the entirely new category of things that become buildable. "Single-use software" and "self-driving software" are his terms for what emerges.

"I think the real unlock is not in efficiency, but in capacity... single use software is going to exist, self-driving software is going to exist. Anything that you want to go do or want to turn into reality — anything that involves using a computer in some form — is ultimately some kind of software, and when all the software drives itself, what do you want to do?" 00:38:19

Why Enterprises Prefer Third-Party Providers Over Labs Directly

Scott Wu gives a structural answer to why enterprises won't consolidate onto a single lab's tooling: model rankings change every six months, organizational retraining costs are enormous, and what enterprises actually need is someone helping them transform their workflows — not just a chatbot thrown over the wall.

"You don't want to teach all of your engineers or your entire team how to use one particular suite and one particular product and then find out, oh, it turns out people aren't using this one anymore because everyone says this other one is better." 00:25:31

2. Contrarian Perspectives

Benchmarks Are Now Meaningless as Capability Signals

Most of the industry treats new model benchmark scores as meaningful. Scott Wu argues the opposite: any well-defined benchmark can now be solved through RL, which means hitting a benchmark proves nothing except that someone defined the task clearly. The hard unsolved problems are the ones that resist benchmarking.

"We're kind of getting to the point where you can solve basically any benchmark. Because what does it mean to have a benchmark? It means you've already defined the task, you've clarified what success or failure looks like... And the truth kind of is, well if you do that, then you can teach the model to go do that." 00:30:03

The Singularity Narrative Is a Distraction, Not a Roadmap

Against the dominant Silicon Valley belief that recursive self-improvement leads to near-term AGI dominance, Scott Wu argues that intelligence alone cannot solve organizational friction, hardware constraints, or tasks with decade-long feedback loops. The bottleneck is not intelligence — it's measurability.

"There's a lot of practical problems out there in the world that are not actually purely intelligence-soluble. There's just like organizational things that have to be done, processes that take time, hardware components — how much GPU compute can you get." 00:29:10

A Single AGI Entity Controlling the World Is the Wrong Future — and Cognition Is Betting Against It

While most AI companies are in a race to capture or be absorbed by a dominant lab, Scott Wu is explicitly building against that outcome. His independence isn't just a business decision — it's a civilizational bet that a plural, distributed AI ecosystem is both achievable and necessary.

"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 us." 00:00:25

The First Value Unlock in Coding AI Is Not Efficiency — Most People Are Still Measuring Wrong

The conventional framing is AI saves developer time. Scott Wu says this is the small opportunity. The large opportunity is the explosion of things that couldn't be built before — and that most organizations haven't even started asking what they would build if software were free.

"The real unlock is not in efficiency, but in capacity. What can we do if we can build so much more?" 00:38:19

3. Companies Identified

Cognition

AI software engineering company, creator of Devin. Mentioned as the primary subject — reached $500M ARR in under three years, raised over $2.5B at a $26B valuation, with customers including Goldman Sachs, Mercedes-Benz, Citi, Dell, Santander, NASA, the US Navy, and the US Army.

"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." 00:00:08

Windsurf

AI IDE company acquired by Cognition. Former team members went to Google in an aqui-hire, leaving behind a 200-person company with strong enterprise go-to-market, platform engineering, and a deployed customer base. Acquired over a weekend.

"Windsurf had an amazing enterprise traction already, had a team that was really had a lot of expertise in going and selling to these folks and figuring out how to actually deliver — a really, really amazing deployed engineering motion." 00:10:22

Brex

Intelligent finance platform combining cards, expenses, and banking with agentic finance built in. Used by Vercel, OpenAI, Anthropic, Granola, and Deepgram. Mentioned as sponsor and as the finance platform Sourcery itself runs on.

"The companies building what's next in AI — from Vercel, OpenAI, Anthropic, Granola, and Deepgram — all made the same call. They all run on Brex." 00:12:27

MongoDB

Developer database platform. Historically significant to Cognition: setting up MongoDB was the first real task Devin autonomously solved, which was the moment Scott Wu's team first believed an AI coding agent was genuinely possible.

"Devin just ran a bunch of stuff for a few minutes and then fixed it... I couldn't sleep that night. Like we were just like, shit, like maybe it does work. Maybe you can just have like an AI engineer buddy that can just do work for you." 00:36:24

Anthropic

AI lab. Mentioned as one of Cognition's model providers in its multi-model neutral approach.

"We work with OpenAI, Anthropic, Google, and so on." 00:20:53

OpenAI

AI lab. Mentioned as one of Cognition's model providers.

"We work with OpenAI, Anthropic, Google, and so on." 00:20:53

AssemblyAI

Voice AI infrastructure company. Builds industry-leading speech-to-text and speech understanding models used by Granola, HeyGen, Ashby, and ClickUp. Mentioned as podcast sponsor.

"AssemblyAI is a voice AI infrastructure layer millions of developers build on." 00:27:34

Cursor

AI IDE competitor. Mentioned in context of a reported $60B acquisition deal with SpaceX, used as a contrast to Cognition's decision to remain independent.

"Recently there's similar news with Cursor... Cognition remains independent." 00:08:26

Cerebras

AI compute company. CEO Andrew Feldman mentioned as part of the broader RAISE Summit series.

"Andrew Feldman from Cerebras." 00:42:38

4. People Identified

Scott Wu

CEO and co-founder of Cognition. Former competitive programmer. Described as one of a founding team where roughly 30 of the first 35-40 employees had previously founded a company. Known for closing the Windsurf acquisition in a single weekend.

"A lot of our team are former founders. Our first 56 people — I think like 30 of us had founded a company before this." 00:17:29

Jeff (Windsurf co-founder)

Co-founder of Windsurf. Participated in the initial acquisition discussion alongside Graham on the Windsurf side.

"The initial discussions were with myself and Russell from our side and then Jeff and Graham from the Windsurf side." 00:17:57

Steven (Cognition co-founder)

Co-founder of Cognition, described as an engineering and product leader. Was one of the original builders of early agent prototypes at Cognition, pre-Devin.

"We had Steven and Walden who are engineering product leaders, my co-founders, going in sitting with their team to understand what's going on with Windsurf." 00:18:25

Walden (Cognition co-founder)

Co-founder of Cognition. The person whose frustrated MongoDB setup attempt in December 2023 became the founding proof-of-concept moment for Devin.

"Walden was setting up MongoDB and then... he had his Devin basically — 'Devin just go, just go try and make it work.' And Devin just ran a bunch of stuff for a few minutes and then fixed it." 00:34:28

Russell (Cognition)

Member of Cognition's leadership team who participated in the initial Windsurf acquisition discussions alongside Scott Wu.

"The initial discussions were with myself and Russell from our side." 00:17:57

Peter Thiel

Legendary investor and co-founder of PayPal. Participated in what Scott Wu describes as Cognition's seed round. Scott Wu challenged Napoleon (a Founders Fund partner) to a poker match to settle valuation terms; Peter Thiel shut the idea down.

"Peter shut it down. He said, 'I don't really think that's the right way to do this.'" 00:41:01

Napoleon (Founders Fund)

Partner at Founders Fund (Peter Thiel's firm). Was "down" for the proposed poker match to settle Cognition's seed valuation terms before Peter Thiel overruled it.

"I put that with Napoleon. Napoleon was kind of down and then Peter shut it down." 00:41:01

Naveen (Mayfield)

Investor at Mayfield. Previously identified, in a separate Sourcery interview, the three fastest-growing categories in AI. Mentioned as having predicted the coding agent category's explosive growth.

"There was an interview I did way back when with Naveen from Mayfield and he laid out the three fastest growing categories." 00:06:02

Justin Fishner-Wolfson

Investor at 137 Ventures. His firm's team were described as big fans of Cognition; submitted a listener question via Christian Garrett about why enterprises prefer third-party providers over labs.

"Christian Garrett — I asked him for some questions, 137 Ventures, they're big fans. We talked about you in their interview we just did with them, with Justin Fishner-Wolfson." 00:25:09

Tony Kim

BlackRock. Mentioned as part of the RAISE Summit interview series.

"Tony Kim from BlackRock." 00:42:38

Andrew Feldman

CEO of Cerebras. Mentioned as part of the RAISE Summit interview series.

"Andrew Feldman from Cerebras." 00:42:38

CJ Desai

Executive at MongoDB. Mentioned as part of the RAISE Summit interview series.

"CJ Desai from MongoDB." 00:42:38

5. Operating Insights

The Weekend Acquisition Playbook: Speed Collapses Uncertainty

Cognition's acquisition of Windsurf succeeded in part because they moved in 72 hours, before the situation could deteriorate further or competitors could organize. The key was immediately decomposing the deal into parallel workstreams — product, go-to-market, and terms — and running them simultaneously across teams that Saturday.

"By Monday morning we had the announcement ready to go... Every single kind of like part of the business that Saturday — we had our technical team, we had Steven and Walden going in sitting with their team to understand what's going on with Windsurf, what do we do to get out Wave 11, what are the things we need to make sure to deliver on time." 00:10:48

Don't Force Integration: Let Usage Pull Products Together

Rather than mandating a product merger timeline post-acquisition, Cognition let actual user behavior and natural feature overlap drive integration. This preserved both products' user bases, avoided forced churn, and let the combined product emerge organically.

"Rather than force it and say over the next 30 days or 60 days we have to integrate these products, it's much more of letting the actual usage and the new features that we're building, the next paradigms, letting those pull the products together." 00:17:02

Measure AI Productivity by Business KPIs, Not Tokens or Spend

Scott Wu explicitly rejects the idea that token usage or compute spend are meaningful productivity proxies. The only valid measure is the business outcome being driven. This has direct implications for how any operator should be evaluating their AI investments.

"What you care about is outcomes and what you're actually delivering. A lot of what we think about is 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." 00:03:24

Hire Founders, Not Just Operators, to Navigate Ambiguous Scaling

Cognition's ability to handle a chaotic, unplanned weekend acquisition with a 35-person team was a direct product of its founding team composition: roughly 30 of the first 35-40 employees had previously founded companies. That instinct to navigate from first principles under pressure is not teachable in a short timeframe.

"A lot of our team are former founders. Our first 56 people — I think like 30 of us had founded a company before this... I think sometimes people describe it as 'I just want to build my product in peace and everything else about building a company is not my cup of tea.' For all of us, the experience of building the company is a lot of the fun of it." 00:17:29

6. Overlooked Insights

The "Abundance Era" Creates an Entirely New Category: Single-Use Software

Scott Wu briefly introduces the concept of "single-use software" and "self-driving software" — software generated on the fly for a specific task and discarded — but the hosts don't follow up on it. This is actually a profound structural shift in the software industry. If software can be generated instantly for any task, the entire premise of software as a durable, maintained product changes. Every workflow becomes programmable on demand. The infrastructure, tooling, and business models built around persistent software (SaaS, licensing, version control) may face a fundamental challenge from this paradigm.

"Single use software is going to exist, self-driving software is going to exist. Anything that you want to go do or want to turn into reality — anything that involves using a computer in some form — is ultimately some kind of software, and when all the software drives itself, what do you want to do, what do you want to create?" 00:39:17

The "Undefendable Benchmark" Problem Is a Massive Signal for Where AI Investment Should Go

Scott Wu's observation that any well-defined task can now be solved by RL — and therefore benchmarks are becoming meaningless — contains a buried corollary that nobody in the conversation surfaces: the most valuable and defensible AI applications are those built around tasks with long time horizons and ambiguous success criteria, precisely because those tasks cannot be easily trained away. Any investor or operator looking for durable AI moats should be looking specifically for problems where it is hard to define what "correct" looks like and where feedback loops are long.

"Not all of the tasks of the world today cleanly fit into that description. There's lots of things that are super amorphous, that don't really have clear definitions, that it's hard to say what counts as success or failure, that have really long time horizons where it takes a long time to find out whether it was the right decision." 00:30:33