· Scott Wu

Scott Wu: Cognition — Building Devin, the AI Software Engineer

Competitiveness as fuel plus ruthless focus on a single domain let a late-entrant startup plant a flag on autonomous AI software engineering (Devin) and refuse billion-dollar acquisitions to build a generational independent business — betting that first-principles reasoning about exponential AI capability beats the pattern-matching that says incumbents win.

aisoftware-engineeringagentsfounder-modecompetitivenessindependenceenterprise-sales0% confidence

Why this is in the corpus

Scott Wu articulates the clearest operator case for AI-agent autonomy on a horizon (seconds to hours to years of unassisted work), the economics of when software gets built at all, and why focus and independence beat resource-rich incumbents — a signal-rich counterpart and productive tension to the Dan Shipper (every agent needs a human) and Mercor/Foody (application-layer defensibility) insights already in the corpus.

Summary for skimmers

Cognition scaled Devin from ~$1M to ~$500M revenue in ~18 months, selling autonomous software engineering to enterprises (Goldman, Mercedes, US government) via a forward-deployed motion that started with the whole company flying to Brazil for Nubank. Wu frames Devin as the future human-computer interface, argues AI task horizons are doubling every few months toward year-long autonomous missions, stays model-neutral ('Switzerland') with a compound model system, and refuses acquisition because building the generational thing is the most ambitious move.

Briefing

What survives the editorial filter

This page should feel like a smart colleague already listened for you and left only the operating logic worth keeping. Not everything said in the episode makes it through.

Trust signal

Direct episode extraction

Best used for

Decision-grade retrieval metadata not yet added for this episode.

Hold lightly

No explicit downgrade reason stored yet for this episode.

Principles

Durable claims that survive beyond the speaker's biography — each with explicit limits, transferability judgment, and evidence.

Principle

Humans cannot intuit exponential curves — a good hunt used to mean days of food, now it means a thousand years

People consistently under-estimate exponential change because human intuition is calibrated for linear payoffs, not compounding ones.

Model the exponential explicitly; your gut will always under-count it.

“a good hunt will bring you, you know, a couple days worth of food or something. But, but, but, but obviously with the kind of exponential curves that we deal with, you know, the equivalent of a good hunt here could be a thousand years worth of food. And, and we don''t have that intuitive signal in our brains to really understand that”Scott Wu

Principle

Software only gets built when it will be used enough times — agents collapse that threshold to one

The economics of software have always required high usage to justify the build cost; agents that generate throwaway software on demand make one-time-use software economical for the first time.

Ask what becomes worth building once the software only has to be used once.

“the, the math only really works out to create software if it''s gonna be used at least like a million times or something. You know, I''m giving a, maybe it''s 10,000 times or whatever.”Scott Wu

Principle

For a mission-driven founder there is no acquisition price — the work is the reward

For a founder whose material needs are trivially met, the mission itself is the payoff, so there is no number that beats continuing to build.

If money past a low bar buys you nothing you want, the mission is the only real offer on the table.

“we would sell if we thought it was the most ambitious thing to do. It''s kind of an oxymoron ... but you know, it''s, it''s like my genuine answer”Scott Wu

“I don''t have like, I don''t have a car ... I like eating sushi. That''s fun. The sushi doesn''t cost that much. You can do that off of an engineer salary as well.”Scott Wu

Principle

Founders are rationally optimistic — they believe they will win with no evidence they should

The defining founder trait is rational optimism — conviction of success in the absence of supporting evidence; the "it''s too late" nihilists are, by definition, not founders.

Optimism without evidence is not naivety — it is the entry requirement for founding.

“Founders are rationally optimistic. ... They believe even there''s no evidence that they should succeed that they will succeed.”David Senra

“Maybe it''s not possible to build a new independent business. ''cause everything else is, you know, it''s like all the ... Labs are gonna do everything”Scott Wu

Principle

Devin''s durable bet: the enduring abstraction is the human-computer interface, not code

The right thing to own is not code but the human-computer interface — the durable function of software engineering that survives every rise in abstraction.

Anchor your product on the function that survives the next abstraction, not the current medium.

“we''re not always, I mean, we''re not gonna be interacting with code for that much longer. ... what is always true probably is that it will be the human Computer Interface. And so what I mean by that is like ... Devin is the way that humans can tell their computers what to do.”Scott Wu

Principle

Measure output, not token spend — the input metric quietly corrupts the incentive

Ranking engineers by AI token consumption optimizes an input; the correct metric is output and good work delivered.

When a new input becomes measurable, resist ranking on it — rank on output.

“people talk about, oh, like, yeah, like we rank our engineers by how many tokens they''re spending. Well let''s, let''s try and rank people by how much output they''re actually producing or how much good work, you know, is actually getting done.”Scott Wu

Principle

In a chosen domain a focused startup out-cares the resource-rich generalist incumbent

A focused startup beats a bigger, richer incumbent inside one domain by simply caring more about that domain than a company doing everything can.

Pick a domain you can out-care everyone in, then narrow relentlessly.

“I love the quote from, from Daniel Eck in Spotify, right? ... he was like, you know, I can give you all the other reasons, but the truth is we''re just gonna care way more about music than they are.”Scott Wu

“we are, we are just gonna care so much more about like, what does it look like to build software end to end at Goldman Sachs or at like Mercedes-Benz”Scott Wu

Principle

In inflection periods, reason from first principles — pattern-matching fails exactly when it matters

Predicting the future by extrapolating history works almost always and fails precisely in the rare moments that matter most, when you must switch to first-principles reasoning.

At a real inflection point, distrust the base rate and reason from primitives.

“in these particular periods where things that actually move and, and they''re real things that are different, those are the, the 1% of times where it truly is different. Now you just kind of, you know, rather than any kind of pattern matching, like what, what really matters is just thinking about things from first principles”Scott Wu

Principle

A startup has no right to exist — you win only by planting a flag on the future and running at it

Startups have no inherent right to win against incumbents; the sole path is a high-conviction stake on a specific future, executed with total commitment.

Your only edge as a startup is conviction on a future the incumbents won't commit to.

“when you start a company you kind of have nothing like you, you have no right to exist is maybe one way to put it.”Scott Wu

“the reason that you are sometimes able to anyway is if you really like, you know, plan your flag on the ground and, and, and put a stake into like what you think the future is and you run like, held towards that.”Scott Wu

Principle

Loss must hurt more than winning feels good — but not enough to make you stop

The productive competitive temperament weights loss more heavily than winning, but keeps that weight below the level that would make you quit.

Channel loss-aversion as fuel, not as a brake.

“losing feels way worse than winning feels good, but not by enough that it makes me want to stop trying. If that makes sense.”Scott Wu

Principle

"Everyone works toward the known end-state" is an uncreative model of a dynamic world

The belief that the best-resourced player inevitably wins assumes a static, single-goal world; the reality of millions of shifting problems guarantees open niches for focused startups.

Whenever someone says the incumbent will just do it, look for the niche the static model misses.

“if, if there was like one thing that, okay, here''s what we all know is gonna be the end state future and everybody''s just working toward it and whoever has the most resources toward it wins. ... But if there are millions of problems out there to solve, there''s lots of different things. The world is dynamic. Things change all the time.”Scott Wu

Principle

Being model-neutral ("Switzerland") aligns your incentives with the customer''s ROI

Positioning as a neutral orchestration layer above the model providers aligns the vendor with the customer''s cost-efficiency instead of with any single lab''s consumption.

Sit above your suppliers so your incentive is the customer''s ROI, not their spend.

“we like being Switzerland. Exactly. And so I think it''s like an important thing of, of, you know, we are just as incentivized as they are to figure out how to make their token spend efficient”Scott Wu

“we''re not incentivized for them to spend more on the models either. ... we''re incentivized for them to get value and to get output out of it.”Scott Wu

Frameworks

Reusable systems and operating models — including when they help and when they break.

Framework

The compound model system — route each subtask to the best model on the quality/cost/speed frontier

Rather than binding a product to one model, decompose each task into subtasks and dynamically route each to the model best on the quality/cost/speed frontier.

The system treats models as interchangeable arsenal components selected per-subtask, which both optimizes cost and keeps the orchestrator neutral across labs.

Use when: Multi-step agent products where subtasks vary widely in difficulty and cost sensitivity.

Skip when: Single-shot, uniform-difficulty tasks where the routing overhead exceeds the savings.

“Devin is purposely meant to be a, you know, a compound model system. ... it''s like some tasks are these really crazy hard ones where you wanna actually use max thinking ... And then many other tasks are, you know, boilerplate enough or repetitive enough that what you care about is just getting it done really fast.”Scott Wu

“Devin can use any of the different models that has, its in its arsenal, which include all of these models from Anthropic, OpenAI, Google, et cetera, but, but also our own models ... And it will, you know, dynamically go and choose these models for these tasks.”Scott Wu

Framework

The agent task-horizon ladder — seconds are commands, hours are tasks, years are permission

Agent autonomy scales through qualitatively different interaction modes — command (seconds), task (hours), mission/permission (years) — as unassisted work-length grows.

The ladder implies the highest-value use of long-horizon agents is granting permission to pursue what you care about, with a manager-AI selecting the missions.

Use when: Designing how much scope and autonomy to hand an agent given its current capability horizon.

Skip when: Assuming a short-horizon agent can be trusted with an open-ended mission it will fail — match the mode to the capability.

“when you''re talking about seconds, you''re literally talking about just like a specific command, right? When you''re talking about hours, you''re talking about giving it a task and having it do the task. ... What you''re talking about is like, you''re giving the agent permission basically”Scott Wu

Framework

The abstraction ladder of programming — predict software''s future by how humans instruct computers

Software history is a monotonic climb in the abstraction at which humans instruct computers; forecasting the next rung tells you where to build.

The framework reframes "coding agent" as merely the current rung; the durable position is always the top of the ladder where humans express intent.

Use when: Deciding where on the human-computer interface stack to build a durable AI product.

Skip when: Short-horizon tooling decisions where the current abstraction layer is stable and won''t move within the product''s lifetime.

“there, there was a time where programming was like using the vacuum tubes ... or it would''ve been like, you know, filling out the punch cards ... or writing, you know, the, like putting down assembly, you know, or writing in basic or something ... So, so we''ve gone through a lot of generations already is my point.”Scott Wu

“at this point, you don''t need to know ... Python or Java or something like that in order to build your own software ... you can just say, Hey, here''s what I want.”Scott Wu

Signals

What appears to be shifting, for whom it matters, and what happens if you ignore it.

Signal

Enterprise AI-coding adoption is compressing from a 12-18 month cycle to ~3 months

The typical 12-18 month enterprise software adoption cycle is compressing to roughly three months for AI coding tools.

Sell to the priority, not the procurement queue, to unlock the 3-month deployment.

“That''s like a, usually for, for typical companies is like a 12 to 18 month cycle. The thing that we do naturally is ... we just work with them to figure out how we go and do that as fast as, as humanly possible.”Scott Wu

“We try to get deployed with folks, you know, within like three months.”Scott Wu

Signal

Agent unassisted-work-time is doubling every few months — from ~20 seconds to hours

The unassisted-work horizon of AI agents is doubling every few months and has already reached hours-long tasks.

Assume the agent work-horizon keeps doubling; do not architect around today''s task length.

“the famous like METR report, which was saying, you know, a couple years ago AI would ... do about 10 to 20 seconds worth of human work without interruption. ... And that, that''s just doubled every, you know, every couple months basically. And now we''re talking about like hours of work.”Scott Wu

Signal

Within ~5 years agents will do most of what a full human-computer interface requires

Wu forecasts most of the general tell-your-computer-what-you-want capability will be solved within about five years.

Bet on the natural-language interface, not the current coding-task surface.

“I honestly, I mean, I think we''ll have solved most of that over the next five years or so. And in AI terms, five years is, is like a century”Scott Wu

Signal

Agents will progress from hours to days to months to a full year of unassisted work

Agent autonomy will extend to months and eventually a full year of unassisted work, turning delegation into mission-assignment.

Prepare to direct fleets of long-horizon agents by getting good at picking missions.

“why can''t that be days? Or why can''t that be weeks or months of work? And then what does the world look like if everybody has an agent that can just do months of work for them at a time?”Scott Wu

“what happens when they can work for a year? Yeah. Unassisted.”David Senra

Signal

Software is eating the world with "a couple orders of magnitude to go"

Software penetration still has a couple orders of magnitude to grow, and every Fortune 500 is now effectively a software company.

Size the market as all enterprise engineering plus newly-economical single-use software.

“software is eating the world is the, is the famous line. I think it''s very much true. It''s still like, it''s still got a couple order of magnitudes to go.”Scott Wu

“all the Fortune 500 or all the biggest companies in the world are software companies in 2026”Scott Wu

Opportunities

Only included where there is a buyer, a real wedge, and a plausible revenue path — not vague idea theater.

Opportunity

Single-use software: automate the one-off white-collar tasks never worth coding before

A wide-open market exists to auto-generate single-use software for the one-off white-collar tasks that were never worth a dev team.

Look for tasks done once and never again — they are now addressable by on-demand software.

“all the white collar work that we talk about today ... wake up in the morning, all right, I''m gonna go look through these like 15 LinkedIn profiles ... or I''m gonna go fill out these forms, or I''m gonna do this data analysis and put this Excel sheet together ... All of these things are things that could be done with software. It''s just, it obviously doesn''t make sense to, to hire a whole team of people to go make you that piece of software which you''re gonna use one time and never again.”Scott Wu

Opportunity

Ten more generations of product surfaces for how software gets built

The way software gets built will pass through ~10 more product-experience generations, each an opportunity to build the next interface.

Treat the current agent UI as one rung; the next nine are open territory.

“I think we have 10 more generations of these different product experiences to come. Right. And like building those and doing those is like, that is what innovation is.”Scott Wu

Lessons still worth keeping

Useful takeaways that did not fully clear the bar for durable principle status.

Lesson

Cognition flew the entire company to Brazil to win Nubank — its first real success

Cognition''s first enterprise win came from deploying the entire company on-site at Nubank to hand-build a custom Devin for a large code migration.

Concentrate the whole org on one customer''s reality to manufacture a first success when the tech is not yet general.

“Our first success ended up being with a company called newbank, biggest bank in Brazil ... And the use case was like one of these big migrations and we had kind of a custom Devin that was like extremely, extremely optimized for doing that.”Scott Wu

“our entire team was the Ford deployed team. Like we all flew to Brazil. ... the whole team was there. We were sitting there with their engineers understanding, okay, so this is what you do.”Scott Wu

Lesson

The viral launch had no product, no revenue, no customers — early pilots mostly failed

Cognition''s viral moment was a demo, not a product; the pilots that followed mostly failed, and that failure taught them where the real PMF was.

The honest-failure phase is a repeatable precursor to the scoped first use case; over-promising during it is the avoidable error.

“it, it wasn''t even a product at the time, to be honest. I mean, it was more just like a prototype or like a demo of what was possible.”Scott Wu

“perhaps unsurprisingly, I mean they were all just like failing, which is kind of natural. ... but it was certainly not ready to work on like a actual company''s like real code base”Scott Wu

Lesson

The MongoDB moment: the first real task Devin completed kept Wu up all night

One successful autonomous run (Devin fixing a MongoDB setup) was the proof point that convinced the founders the paradigm was inevitable.

Founder conviction can anchor to one existence proof rather than aggregate metrics — a hallmark of believers in exponential tech.

“The first task that Devin did was it like set up MongoDB for us ... at some point we were just like, okay, dev, just, just try to go fix it. Just go run the commands. Like do whatever you need to go do it. And then it worked”Scott Wu

“I remember the first time that it did like a real task. Like I could not sleep that night.”Scott Wu

Lesson

Devin launched at 13% on SWE-bench — 3-4x the prior best — yet the low absolute number invited hate

Devin''s March 2024 launch at 13% SWE-bench was 3-4x the best-known 3-4%, yet the sub-majority absolute number drove polarized criticism.

Relative-progress and absolute-performance readings diverge sharply for early exponential-tech launches; expect backlash on the absolute number.

“SWE bench? Yeah. It was at like, it was like 13%.”Scott Wu

“At the time the best known was like three or 4% or something like that. But obviously yeah, 13% it still means you fail, you know, 87% of the time.”Scott Wu

The Plays

Try these this week

Verb-first executable actions — each one tied to a stated outcome in the episode.

Fly the entire company to your first enterprise customer to hand-build the product around their workflows

Outcome: Deploy the whole company on-site at the first enterprise customer to build the product against their exact reality.

“our entire team was the Ford deployed team. Like we all flew to Brazil. ... We were sitting there with their engineers understanding, okay, so this is what you do. In that case, this is what you do in that case, and this is what Devin needs to know”
— Scott Wu

weeks to months of co-location per

Before you start

· a working prototype
· a team willing to travel
· a customer with a bounded high-value task

Plant a contrarian flag publicly with a viral demo of real runs — before it is a product

Outcome: Publicly plant your contrarian thesis with a demo of genuine (if cherry-picked) runs to own the category first.

“all of those, like demos that we showed, you know, in, in, in our launch announcement were like actual runs of Devin that we had done ourselves and like run into and been like, holy shit, this is insane.”
— Scott Wu

single launch moment per

Before you start

· a genuine contrarian thesis
· real runs worth showing
· willingness to be criticized

Enter the market on the repetitive, scoped, tight-feedback task the agent can actually do

Outcome: Pick the repetitive-scoped-tight-feedback task (e.g., code migrations) as the agent''s first PMF wedge.

“it should be some of these like really repetitive t ds ones ... but it is like repetitive enough and scoped enough and like on a tight enough feedback loop that you can have an agent do it”
— Scott Wu

initial GTM phase per

Before you start

· honest read of current agent capability
· access to enterprise migration-style work

Compress the enterprise deployment cycle from 12-18 months to ~3 by making it a priority

Outcome: Turn a 12-18 month enterprise cycle into ~3 months by elevating it to priority and pre-clearing security.

“usually for, for typical companies is like a 12 to 18 month cycle. ... We try to get deployed with folks, you know, within like three months.”
— Scott Wu

~3 months vs 12-18 baseline per

Before you start

· private-cloud deployment capability
· security and data-agreement tooling
· an executive sponsor

Stay model-neutral ("Switzerland") and run a compound system routing each subtask to the best model

Outcome: Be neutral across all labs and dynamically route each subtask to the best model on the quality/cost/speed frontier.

“we like being Switzerland. Exactly.”
— Scott Wu

continuous per

Before you start

· multi-model integration
· per-subtask quality evals
· neutral commercial model

Decline billion-dollar acquisition offers to keep the independence needed for a generational mission

Outcome: Turn down acquisition offers when the mission''s ceiling is higher than the price and independence is the constraint that matters.

“we would sell if we thought it was the most ambitious thing to do.”
— Scott Wu

ongoing per

Before you start

· founder alignment on mission over money
· capital independence

Run an incentive-aligned forward-deployed motion — point customers to the right use cases and tell them where NOT to use you

Outcome: Guide customers to the high-ROI use cases and openly tell them where the agent will fail — honesty as an adoption engine.

“here are each of the projects that we think Devin right now can make you 10 times faster on. And we''ll tell you for the ones that, that it''s not. And here''s, here''s what workflows you should use instead”
— Scott Wu

ongoing customer partnership per

Before you start

· a forward-deployed team
· honest capability assessment per use case

Sell enterprise by quantifying ROI against the outsourced alternative

Outcome: Frame the pitch as the ROI delta versus the customer''s specific outsourced alternative, in dollars and months.

“this project which was scooped out for 18 months and was gonna be handed off to, you know, an outsourced contractor and, and was gonna cost you 15 million. Like, let''s just talk about how that you do this all internally with your own team and you do it for 1 million and you get it all done in three months.”
— Scott Wu

per deal cycle per

Before you start

· visibility into the customer''s planned spend
· a scoped first project to prove value

Decision Moments

Actual decisions, real outcomes

Specific decisions narrated in the episode with their outcomes and transferable lessons.

Early 2024, a year-plus after ChatGPT, entering a field with GitHub Copilot, OpenAI, and DeepMind entrenched, with only a prototype (not a product) that succeeded on rare runs.

Did: Publicly launched a viral demo of real Devin runs and planted the contrarian flag that AI would be a coworker, not an autocomplete tool — then refused to retreat to a safer chatbot Q&A product after receiving heavy criticism.Outcome: Being first to plant the "AI coworker" flag became decisive for brand, recruiting, and customer work; the company scaled ~$1M to ~$500M revenue in ~18 months. Wu concedes they were early and "maybe should have done something in the middle."

Planting a contrarian flag first can be worth shipping ahead of the product, but only with honest demos and the nerve not to retreat under criticism.

Part of an emerging decision pattern across multiple episodes

April/May 2024: GPT-4-era agents did not work generally, POCs were failing, and users were pointing Devin at everything from mundane to deep-architecture tasks.

Did: Bet the product on the repetitive-scoped-tight-feedback task class (enterprise migrations / version upgrades) as the first PMF wedge, built a custom hyper-optimized Devin for Nubank by flying the whole team to Brazil, and architected Devin as a neutral compound model system routing subtasks across labs.Outcome: Nubank (largest bank in Brazil) became the first real success on a large migration; enterprise grew to ~75-80% of revenue with customers like Goldman Sachs, Mercedes, and parts of the US government.

Match the wedge to what agents can actually do today (repetitive, scoped, tight-loop), concentrate the whole org on the first customer, and stay model-neutral to align with customer ROI.

Part of an emerging decision pattern across multiple episodes

Dozens of acquisition approaches (only a handful of buyers could afford them), amid AI-era nihilism that independent businesses are no longer possible and high-profile competitor acquisitions.

Did: Declined to sell, holding that Cognition would only sell if selling were the most ambitious move, and positioned the company publicly as the one betting on independence.Outcome: Cognition remained independent and became known as the "Independence" bet in the space; Wu frames the only intolerable outcome as not having pushed as hard as they could.

For a mission-driven founder whose needs are met, no price beats the generational outcome; independence is the constraint that protects the ceiling.

Part of an emerging decision pattern across multiple episodes

Assembling a nine-person founding team and a company culture, most of whom had already founded their own companies and were lifelong competitive programmers.

Did: Deliberately concentrated talent density around competitive-programming "nerds" — building the company as a physical manifestation of a high-density technical social network where, only half-jokingly, "you''re not allowed here if you''re not a nerd."Outcome: A big (nine-person) founding team of prior founders and elite programmers, described as unusually homogeneous in composition and central to Cognition''s identity and generational ambition.

Talent density and cultural concentration can be an intentional design choice, not an accident — hire for a specific high-density profile and let it define the company.

Part of an emerging decision pattern across multiple episodes

Tensions surfaced

Contradictions and trade-offs the episode raises — judgment calls a thoughtful operator has to navigate.

Tension

Ship the flag-planting demo early vs. the fair criticism that it was not a product

Both are true: releasing early was strategically decisive AND the criticism that it wasn''t a real product was fair.

“it wasn''t even a product at the time ... it was more just like a prototype or like a demo of what was possible.”Scott Wu

“maybe we should have done something in the middle. You know, I, I don''t know.”Scott Wu

Tension

AI will automate everything, yet the belief that everyone loses their jobs tomorrow is wrong

Both are true: AI will automate vast amounts of work AND the mass-unemployment-tomorrow reading is wrong — the shift is to humans in "creative mode" directing agents.

“there were some people who were like, oh my God, everybody''s gonna lose all their jobs tomorrow. Which is not what we''ve ever really believed.”Scott Wu

“We''ve been spending all this time living in survival mode as a species ... and now we''re gonna be living in creative mode.”Scott Wu

Tension

Customer of the labs and competitor with them — the frenemy resolved by focus and neutrality

Both are true: the labs are Cognition''s suppliers AND its competitors — the relationship is positive-sum, not zero-sum.

“How Do you think about this? Like you''re customer of them but also competitor with”David Senra

“in, in practice there''s a lot of positive some work for us to do together. ... software''s the only thing we, that we care about”Scott Wu

Corpus connection

Where this episode fits for retrieval

What kinds of decisions this briefing is best pulled into.

Primary decisions

• strategic-bet
• product
• hire