Devin's Scott Wu on Building an AI Software Engineer and Turning Coders into Architects

Wu describes Devin as a fully autonomous software engineer that works on tasks end to end through an asynchronous workflow. You can tag it on an issue in Slack or Linear, and it makes pull requests in your GitHub. It is built to work with engineering teams as a junior engineer.

A year before this conversation, Wu says, people did not really believe agents were possible, and model reasoning was far earlier. He compares Devin's progress to a high school CS student, then a college intern, and now a junior engineer. He prefers the phrase "jagged intelligence", since Devin is much better than a human at some things and much worse at others.

Much of the past year's work went into the product side rather than raw capability: Slack and Linear integrations, an interactive planning phase, and ways to touch up Devin's code. Wu calls the split about 50/50 between capabilities and interface, since people know how to use chatbots but still face a learning curve with agents. Customers range from one- or two-person startups to Fortune 100 companies and public banks.

The founding team has known each other for years, many from competitive programming and math competitions. Wu previously ran Lunchclub for about five years, one co-founder was an early engineer at Scale AI and another at Cursor, and others came from Waymo and other startups. They got together in late 2023 with two convictions.

The first was that high-compute reinforcement learning would be the next paradigm after the imitation learning behind the first ChatGPT, in which models do work on a task and are evaluated on whether it was correct. Code was the natural domain because the team are programmers and because running code gives an automated feedback loop. The second was that products would shift from text completion, like GitHub Copilot, to agents that make decisions, take feedback and iterate over multiple steps.

The company started as a hackathon-style project in an Airbnb around Thanksgiving 2023, first applying an agentic loop to contest programming problems. Another hacker house produced the early ideas for a software engineering agent, and even the ability to talk to Devin mid-task had to be invented. The timeline Wu gives is a start in November 2023, a company in early 2024, a March launch, self-serve availability in December and a 2.0 release a few weeks before the interview, with eight pivots within coding agents along the way.

Asked why Devin has a name and personality unlike most AI apps, Wu says it is a decision the team is proud of. The most useful way to explain the experience is "your junior buddy" that you hand things off to. Early users often faced a blank screen or asked for a huge re-architecture of an entire code base.

The team learned to steer people toward setting up the repository first, giving Devin a few small tasks to learn the code base, and making sure it has its own virtual machine to run tests, linters and CI. The best team workflow is running multiple Devins asynchronously and jumping in only to give feedback or steer the plan. The name, Wu says, is an attempt to capture that of an autonomous entity you teach and learn with over time.

Wu says programming will only become more important as AI gets more powerful, because the discipline is fundamentally about telling a computer what to do. He does not expect computers to take over everything soon. Meanwhile, working with Devin is like having a team of junior buddies, and each engineer on his team has their own.

Devin merges several hundred pull requests a month into the Devin code bases, from a team of only 15 engineers. Sometimes Devin does a task fully autonomously; other times the engineer steps in for the final 10 or 20%, such as scoping, architecture or testing the front end. About a quarter of PRs are now from Devin, and Wu expects it to be more than half by the end of the year.

Wu says the soul of software engineering is defining the problem and thinking through architecture and details, which is the most fun part but roughly 10% of an average engineer's time. The other 90% is Kubernetes errors, bug reports, migrations and upgrades. Devin is meant to let engineers go from bricklayer to architect while keeping the human in control of specification.

On whether people should still learn to code, Wu says absolutely yes. Computer science mostly teaches how to break problems down logically and the model of a computer, such as databases, garbage collection and networking, rather than syntax. Even as Python and now English-like instructions abstract details away, understanding the layers beneath helps build good systems.

The skills that grow in value are understanding complex systems, working in the context of the whole company, and making key product decisions and trade-offs. Wu predicts there will be far more programmers and engineers a few years from now, even if the form factor changes. He calls software the shining example of Jevons Paradox.

Explaining the paradox, Wu says that as the price of something falls, total spending on it can rise. A zero-sum view says 10 times faster engineering means 10 times fewer engineers, but he expects society to build more than 10 times as much code. People, he says, will build many more personalized experiences.

Most people on Wu's team work with up to five Devins at a time, starting and stopping them as needed. An engineer thinks through the five things to get done that day and assigns each to a Devin. Wu says it took time to feel intuitive, but the engineer's role is to supply expertise where it counts, such as defining exactly what is being built or steering complex changes.

Host Lenny Rachitsky notes the team has about five times as many Devins as engineers and sees it as a glimpse of how most companies may operate. Wu says engineers on his team already spend less time on boilerplate and more on questions like how to make Devin better and what the right interface and flow are. He expects more work to shift to asynchronous flows because real-world engineering, from reproducing a bug to checking Datadog logs and rerunning, takes real time.

In the first demo Wu asks Devin, working on its own code base, to feature Lenny's newsletter on the Devin website. Devin starts immediately, explores the files, proposes a new sidebar component and asks questions, such as whether the button should open in a new tab. He can answer at any point, and Devin produces a pull request, which later shows the newsletter on the home page.

Wu also shows a research task he gave Devin that morning: research the host and build a quiz website so he could check his facts before the podcast. Devin browsed sources, found the newsletter and built the app, and the two then played the quiz live. A third example imports the DeepSeek repository, where Devin builds a wiki with architecture diagrams and answers questions about the code base.

Wu says users find the wiki useful for scoping tasks and for onboarding new engineers who may feel awkward asking basic questions. Rachitsky compares it to a staff engineer's knowledge of the code, and Wu says language models are strong at retrieval and processing lots of code. The last demo is a Linear integration: add a Devin label and it posts its thoughts and a confidence level, and product managers like Claire Vo use it to scope tasks and ask data questions.

Wu says to give Devin tasks, not problems: quick front-end requests, bug fixes, tests and documentation. The best tasks are easy to verify and quick to iterate on, such as previewing a link, and Devin will sometimes log into Devin itself to check its work. Bigger asks are possible, but you should expect to steer more.

Rachitsky notes this resembles synthetic data and reinforcement learning, where clear yes or no answers make the best signal. Asked about internal debates, Wu says one is how opinionated to be. Most features serve pull requests and engineering teams, but people have used Devin for odd things like ordering DoorDash, and he notes the usual advice to focus on a niche sits awkwardly with general-purpose generative AI.

The other debate is whether Devin should be one comprehensive experience or a suite of tools. The team now sees a suite, including Devin Search, Devin Wiki and Linear ticket scoping, with the core agent at the center. Users like control, such as asking questions about a repo without starting a task, so the suite serves both capability and UX.

Asked about competitors including OpenAI, Anthropic, Cursor, Windsurf and Replit, Wu calls them all incredible teams and says there is much to do. He argues AI will speed up every niche of the software lifecycle, from incident response to logging to billing, by an order of magnitude. Cognition's lens has been autonomous coding agents, and he says there are perhaps 20 generations of agent product to come.

He imagines a later generation where you never look at code, only at your product, saying a button should be rounder or asking for a new database table. The team has been laser-focused on agentic coding, which extends to its usage-based pricing with ACUs and the places you can talk to Devin. Wu expects to stay in that space for the next five or ten years.

On defensibility, Wu says stickiness is a better frame than moats, since he sees no hard barrier keeping others out of any layer of AI. Devin becomes stickier as it builds a representation of your code base, stack and process, like an engineer who has been at a company for years. It also has a multiplayer aspect: engineers teach it in Slack, comment on its PRs and use it to onboard new colleagues.

Rachitsky compares this with a Cursor view that staying the best matters most, with stickiness as an extra. Wu adds that software engineering has a clear tie to value and a clear next level for a long while, perhaps eventually something like building all of YouTube out of the box. Developers, he says, are willing to learn new experiences and put in effort for higher-quality results.

Asked what unlocked Devin, Wu says no single model shift made a night-and-day difference, though every new model moves the curve, and Cognition works with the research teams at foundation labs. His hot take is that base intelligence is already roughly there. The team's effort goes into teaching the model the idiosyncrasies of real-world engineering, such as using Datadog, diagnosing errors and making a GitHub PR.

Wu contrasts AI with earlier tech revolutions like the PC, internet and mobile phone, which were paced by hardware distribution and grew steadily over years. AI has no such weight and can grow explosively. He says we are past the inflection point in AI code, and an engineer not using AI to write code is falling behind.

For adoption, he says a few enthusiasts typically set up Devin, give it repos and teach it linting and CI, then give it initial tasks. Others notice the stream of PRs and join, and by then Devin already knows the repositories. His one-line tip for new users is to treat Devin like a new junior engineer, start with the easier tickets, and scale up.

Rachitsky worries that engineers may become managers of many junior engineers, reviewing PRs all day. Wu says bricklayer versus architect is closer to the experience than manager. He thinks management is disliked for its context, ownership and emotional aspects, whereas working with Devin is an interface for handing off tasks, comparable to the shift when Python was invented.

The key is finding the right level of abstraction and workflows where Devin takes a first pass, you merge it or give feedback, and Devin stays part of your flow without you losing control. Devin can already start other Devins through the API and often chunks big tasks that way, though it needs credentials and is not on by default. Wu adds that humans couple context to a single thread, while an agent can pursue many lines of exploration while sharing all context.

Asked for something counterintuitive, Wu says that although he learned thousands of details building companies, success often comes from doing three to five familiar things much harder than expected. Moving fast meant hackathons in November and December, an official start in January, prototypes with users in February, a launch in March and first customers in April. He notes that 18 of the team of 26 or 27 had founded companies before.

On hiring, he says you should fight to all ends for the people you want. He tells of an MIT junior who interviewed better than almost any full-time candidate; the team flew to North Carolina, had dinner with his parents and worked out an arrangement where he works full time and comes in for required classes to graduate. For another strong candidate they helped hand-write rejection notes to other companies to preserve relationships.

His summary of the basics is to hire the best people, move fast, build something people want, stay close to customers and think about where things are going rather than where they are. In AI, he says, that means thinking about next week as well as ten years out. His closing note is optimism: he sees everyone able to multiply themselves and is not worried about running out of things to do.

In the lightning round, Wu recommends *The Power Law* by Sebastian Mallaby and *The Great Gatsby*, and likes the Aura picture frame and AirPods. His motto is to give everything while not tying your personal worth to success or failure. The name Devin came from merging the virtual developers each founder built, such as DevWalden and DevSteven, into one universal dev, and the logo exists as both hexagons and an otter.