LLenny's Podcast
← All episodes
Scott Wu (CEO and co-founder of Cognition)04 May 2025

Inside Devin: The world’s first autonomous AI engineer that's set to write 50% of its company’s code by end of year

7Frameworks
15Insights

Frameworks in this episode

Leadership3 steps

Champion-Led Agent Adoption

Roll out new AI tooling by letting a few excited early adopters pave the way for the team.

Read →
Entrepreneurship4 steps

Do The Five Fundamentals Harder

Startup success isn't secret knowledge — it's executing the well-known basics far past normal effort.

Read →
Mindset3 steps

From Brick Layer To Architect

Reclaim the 10% of high-leverage architecture work by delegating the 90% of implementation to AI.

Read →
Strategy4 steps

Stickiness Over Moats

Stop chasing barriers competitors can't cross; build accumulating value they can't easily replace.

Read →
Productivity4 steps

Tasks, Not Problems

Delegate to AI agents by handing them scoped, verifiable tasks — never open-ended problems.

Read →
Productivity3 steps

The Agent Foothold Onboarding

Onboard an AI agent like a new hire: environment first, easy tasks next, then scale.

Read →
Productivity4 steps

The Parallel Agent Workflow

Run several AI agents at once and insert yourself only where your expertise actually matters.

Read →

Insights & moments

The myth-busts, hot takes, explainers, and tools worth keeping.

Hot Take· 3

Hot Take57:30

Forget Moats, Build Stickiness

On the perennial AI defensibility question, Scott reframes it away from moats and toward stickiness. He doesn't believe any hard barrier prevents competitors from entering, but coding agents accrue real stickiness: Devon learns your codebase and process over time like a long-tenured engineer, plus a 'multiplayer' effect as a whole team teaches it.

  • Moats imply a competitor can't even enter; Scott doesn't think such barriers exist
  • Stickiness is whether you're excited to keep using an experience versus easily switching
  • Devon builds up knowledge of your codebase and process over time, like a five-year veteran
  • A 'multiplayer' aspect compounds value as an entire team teaches and reviews Devon
  • The goal isn't to lock others out but to make Devon more useful the more you use it

“I think it's often less about moes and more about stickiness.”

Scott Wu · 57:30

“there is a lot of just inherent kind of stickiness and learning and buildup over time, which is that as you use Devon and as…”

Scott Wu · 58:00
#moats#defensibility#strategy#ai-products
Hot Take1:02:30

Base Model Intelligence Is 'Basically Already There'

Scott's hot take on the tech behind Devon: there was no single step-function model release that made Devon night-and-day better, and in terms of base intelligence, we're basically already there. The work now is less about raising a model's IQ and more about teaching it the idiosyncrasies of real-world engineering.

  • No single base-model shift made a night-and-day difference to Devon
  • A new model roughly every week has steadily moved the curve
  • Cognition doesn't pre-train its own models; it builds on the foundation labs
  • The real work is teaching the model real-world idiosyncrasies (e.g. using Datadog, diagnosing errors, making GitHub PRs)
  • It's about mirroring the complexity of the real world, not reaching a higher fundamental problem-solving level

“my my hot take here which I would give is I think the I think in terms of base intelligence we're honestly basically already there”

Scott Wu · 1:02:30

“teaching the model to mirror the complexity of the real world I would say rather than rather than kind of like getting it to some…”

Scott Wu · 1:04:00
#ai-models#reasoning#devon#hot-take
Hot Take1:27:00

Give It Everything, But Don't Tie Your Worth to Wins and Losses

Scott's life motto, which he applies constantly in startups, holds two seemingly contradictory truths at once: be intensely focused and driven to maximize your potential, yet don't let your personal emotion get tied up in success or failure. He argues detaching your self-worth from outcomes is not only healthier but actually makes you more successful.

  • Many proverbs contradict each other yet both are true; the skill is understanding why
  • Be focused and driven to maximize potential, but don't tie your emotion to success or failure
  • Startups always have ups and downs, even the most successful companies
  • Detaching self-worth from outcomes lets you give your best without the weight
  • Scott links it to a Buddhist idea of not clinging to a particular outcome

“it's also very important to not let your own personal emotion get tied up in your success or failure”

Scott Wu · 1:27:30

“it also actually makes you more successful.”

Scott Wu · 1:29:00
#mindset#founder-psychology#resilience#life-motto

Explainer· 4

Explainer07:30

How Devon Grew From a High-School Coder to a Junior Engineer

Scott Wu describes Devon's capability progression over its first year as a rough seniority ladder: high-school CS student to college intern to junior engineer. He stresses these are only loose guidelines and prefers the phrase 'jagged intelligence' because an agent can be far better than a human at some things and far worse at others.

  • A year of progress moved Devon from a high-school CS student level to college intern to junior engineer
  • 'Jagged intelligence' captures that the agent beats humans at some tasks and trails them badly at others
  • Much of the year's learning was about how humans should work and interact with agents, not just raw capability
  • Roughly half the improvement came from product interface and tooling, half from model capability

“Sometimes we kind of say, well, when we got started it was kind of like a a high school CS student and then as time…”

Scott Wu · 07:30

“I I I really like the phrase jagged intelligence for example.”

Scott Wu · 07:30
#ai-agents#devon#capabilities#jagged-intelligence
Explainer17:30

Why Cognition Made Devon a Named 'Person' Instead of a Tool

Unlike other AI coding products, Devon has a name and a personality. Scott explains this was a deliberate, proud decision: framing Devon as 'your junior buddy' helps users understand the hand-off workflow, and the name is an attempt to capture the soul of an autonomous entity you teach and learn with over time.

  • Devon is deliberately positioned as a named junior engineer you hand tasks to
  • The 'junior buddy' framing shapes onboarding: start with one-pointer tasks, set up the repo and CI first
  • New users often stall on a blank screen or over-scope a full re-architecture on day one
  • The name is meant to capture the soul of an autonomous entity you teach over time

“And so in many ways I think it's I think Devon as a name really is is is our attempt to kind of capture the…”

Scott Wu · 19:00
#product-design#devon#personality#onboarding
Explainer29:00

Jevons Paradox: Why Cheaper Coding Means More Programmers

Scott explains Jevons paradox, that when the price of something falls, total spend on it can actually rise, and applies it to software. Rather than needing 10x fewer engineers because AI makes coding 10x faster, he argues we'll build far more than 10x as much code, because demand for software has always been effectively unbounded.

  • Jevons paradox: as a good's price drops, total spend on it can still go up
  • The zero-sum view says 10x faster engineering means 10x fewer engineers needed
  • Scott argues the opposite: we'll build more than 10x as much code
  • Software is the shining example of Jevons paradox; society keeps finding more to build
  • Expect way more programmers and engineers a few years from now, not fewer

“Jevans paradox just says that as the the price of something goes down it can still be the case that uh the total spend on…”

Scott Wu · 29:00

“we're going to be 10 times faster at software engineering And so it means that we're going to need 10 times fewer software engineers, right?”

Scott Wu · 29:30
#jevons-paradox#economics#future-of-work#software
Explainer1:04:00

Why AI Spreads Faster Than the Internet or Mobile

Scott contrasts AI with the last 50 years of tech revolutions, PCs, the internet, mobile, which all had a hardware distribution component that made adoption grow steadily over many years. AI has no such hardware bottleneck, so it can grow explosively; he argues we're already past the inflection point in AI code.

  • PC, internet and mobile revolutions all grew steadily due to hardware distribution
  • The internet went from a few universities to the whole world, but it took years
  • AI has no hardware-distribution weight slowing it, so the space grows exponentially
  • We're firmly past the inflection point in AI code; engineers not using it are falling behind
  • Scott calls AI the biggest technology shift of our lives

“AI is going to be the the the biggest technology shift of our lives”

Scott Wu · 1:04:00

“there's no weight on hardware distribution that that is causing that. And it means that that it it means that the space is just growing…”

Scott Wu · 1:06:00
#ai-adoption#technology-shifts#distribution#macro

Story· 4

Story14:00

Eight Pivots and a Hacker House: The Origin of Devon

Cognition began in November 2023 as a Thanksgiving-time hackathon in an Airbnb, initially building agents to solve contest-programming problems. Scott recounts going 'from hacker house to hacker house,' pivoting roughly eight times within coding agents before landing on a full software-engineering agent rather than just a coding tool.

  • Started November 2023 as a hackathon project, not a company
  • First build target was contest-programming problems using an agentic loop
  • The team pivoted about eight times within coding agents over 18 months
  • Became a company early 2024 and launched Devon in March 2024
  • Talking to Devon was itself an invented idea; originally you just handed off a task and got finished code back

“But even with that, it's I feel like we've pivoted like eight times or something within coding agents, you know, over the last year and…”

Scott Wu · 14:00

“the the the story of the whole company for us in some sense has been going from hacker house to hacker house”

Scott Wu · 15:00
#startup#origin-story#pivots#cognition
Story21:00

15 Engineers, 5 Devons Each, a Quarter of All PRs

Scott shares how Cognition uses Devon to build Devon: a team of about 15 engineers, each running up to five Devons at once. Devon merges several hundred PRs into production monthly and currently accounts for roughly a quarter of all pull requests, which they expect to exceed 50% by year's end.

  • Engineering team is ~15 people; most engineers run up to five Devons at once
  • Devon merges several hundred PRs into production every month
  • About a quarter of all PRs are currently Devon's, up exponentially over six months
  • Expected to be more than half of PRs by the end of the year
  • Sometimes Devon does a task 100% autonomously; often the human steers the last 10-20%

“we use a ton of Devon when we're building Devon. And so Devon merges like several hundred pull requests into production in the Devon code…”

Scott Wu · 21:00

“it's in the neighborhood of of a quarter or so of all of our”

Scott Wu · 22:00
#ai-agents#productivity#devon#engineering-teams
Story1:15:00

The Counterintuitive Lesson: Do the 3-5 Startup Basics Harder Than Anyone

Scott's counterintuitive takeaway is that building a company well often comes down to doing the same three-to-five startup cliches, move fast, hire great people, build what people want, far more intensely than you'd expect. He illustrates 'fighting to hire great people' with flying to North Carolina to have dinner with a young candidate's parents, and hand-writing another candidate's rejection letters to other companies.

  • Of 26-27 people on the team, about 18 had started their own company before
  • The real edge is doing the 3-5 basics more intensely than you think possible
  • Cognition ran hackathons in Nov and Dec, incorporated in Jan, launched in March, first customers in April
  • To land an MIT junior, they flew to North Carolina and had dinner with his parents to make school work
  • For another candidate, they hand-wrote his rejection responses to other companies to preserve relationships

“building companies well sometimes just comes down to doing those three to five things just even more than you could possibly expect.”

Scott Wu · 1:16:00

“hiring great people is one thing but truly just never never giving up”

Scott Wu · 1:19:00
#startups#hiring#founder-lessons#culture
Story1:29:30

Where the Name 'Devon' Came From

Scott shares the origin of the name: early on, each founder built their own 'dev' version, Walden made Dev Walden and Steven made Dev Steven, and when they consolidated everything into one universal dev, it became Devon. The bigger debate was over Devon's image; they still keep both the hexagons and a little otter with a laptop.

  • The name stuck very early while working on coding agents
  • Founders each built their own version: Dev Walden, Dev Steven
  • Consolidating them into one universal 'dev' produced Devon
  • The harder decision was Devon's visual identity
  • They kept both the hexagon logo and an otter with a laptop

“Walden made a virtual you know developer version of him which was called Dev Walden and then Steven made one of him which was called…”

Scott Wu · 1:30:00

“somehow we still we still have both the the hexagons and the otter.”

Scott Wu · 1:30:30
#branding#origin-story#devon#naming

Q&A· 1

Q&A25:30

Should You Still Learn to Code? Scott Wu Says Absolutely Yes

Asked the now-classic question of whether people should still learn to code, Scott answers an emphatic yes. His reasoning: computer science teaches you to logically break down problems and to understand the abstractions beneath the surface, from databases to garbage collection to TCP/IP, which stays essential even as AI abstracts more away.

  • Learning to code teaches logical problem breakdown, not just language syntax
  • Understanding abstractions (databases, garbage collection, networking) lets you build good systems
  • Python already lets you 'explain in English' compared to 50 years ago, and AI extends that further
  • For a long time you'll still need to think precisely about details and peel back abstractions

“First of all, the question of, you know, whether you should still learn to code, my answer would be absolutely yes.”

Scott Wu · 25:30

“most of what you're learning really is about the ability to logically break down problems for number one.”

Scott Wu · 26:00
#learn-to-code#careers#computer-science#education

Takeaway· 3

Takeaway24:30

From Brick Layer to Architect: Where Engineers Spend Their Time

Scott argues the fun, high-value part of software engineering, defining the problem and designing the solution, is only about 10% of an engineer's time, while 90% is implementation drudgery like debugging Kubernetes errors and migrations. Devon is built to shift engineers from brick layer to architect while keeping the human in control of the full specification.

  • The most valuable work, defining problems and architecture, is ~10% of an engineer's time
  • The other ~90% is implementation: debugging, migrations, upgrades, bug reports
  • Devon is designed to move engineers from brick layer to architect
  • The human stays in control and does the full specification; Devon multiplies the magnitude

“one of the the the ways that we've kind of thought about Devon uh in building Devon is is is really allowing engineers to to…”

Scott Wu · 24:30

“I I think at the same time that's probably, you know, in the neighborhood of of 10% of the average software engineer's time, right?”

Scott Wu · 24:00
#software-engineering#future-of-work#architecture#devon
Takeaway47:30

Give Devon Tasks, Not Problems: Where Agents Excel and Fail

Scott's rule of thumb for what Devon is good at: give it well-defined tasks, not open-ended problems. Devon shines on quick front-end changes, bug fixes, and adding tests or docs, anything easy to verify and test. Bigger, fuzzier asks are possible but require much more steering.

  • Devon is best on well-defined tasks; 'give Devon tasks, not problems'
  • Ideal work is easy to verify and easy to test
  • Good examples: front-end feature requests, bug fixes, adding testing and documentation
  • Bigger projects work too, but expect to steer Devon much more
  • Parallels how easy-to-verify answers benefit synthetic data and reinforcement learning

“I think Devon is best when it is working on tasks that are well defined. You know, it's one way to put it is you…”

Scott Wu · 47:30

“you you generally want something that is kind of like e easy to verify and easy to test is the main thing.”

Scott Wu · 48:30
#ai-agents#best-practices#devon#workflow
Takeaway1:07:30

How Devon Adoption Spreads Inside a Company

Scott describes the pattern for getting Devon adopted at a company: a few enthusiastic early adopters do the setup, teaching Devon the repos, lint and CI, and start it on foothold tasks. As teammates watch that 'new hire' knock out PRs, they sign on, and by then Devon already knows the repositories they work in.

  • A few excited early adopters put in the setup investment first
  • They teach Devon the repos, how to run lint and CI, and give it initial foothold tasks
  • Teammates notice the 'Devon person' knocking out PRs and get accounts
  • New adopters benefit because Devon already knows their repositories
  • Early adopters pave the way for everyone else on the team

“there will be a few folks at the team who are really excited uh and want to try out the new thing”

Scott Wu · 1:07:30

“the early adopters themselves can really pave the way I think for for everyone else on the team.”

Scott Wu · 1:08:30
#adoption#change-management#devon#teams