Inside Devin: The world’s first autonomous AI engineer that's set to write 50% of its company’s code by end of year | Scott Wu (CEO and co-founder of Cognition)

May 4, 2025 Episode Page ↗
Overview

Scott Wu, co-founder & CEO of Cognition, discusses Devin, their autonomous AI software engineer. He shares how their 15-engineer team uses multiple Devins to build Devin, expecting over 50% of PRs by AI, and how AI will transform engineers from coders to architects.

At a Glance
14 Insights
1h 32m Duration
18 Topics
6 Concepts

Deep Dive Analysis

Introduction to Devin: Autonomous AI Software Engineer

Devin's Evolution: From High School CS Student to Junior Engineer

Devin's Origin Story and Founding Team's AI Journey

Shift from Text Completion to Autonomous Agents

Cognition's Internal Use of Devin for Engineering Workflow

Impact of AI on Software Engineering and Programmer Roles

Jevon's Paradox and Increased Demand for Software

Live Demo: Devin Modifies a Web Application

Live Demo: Devin Researches and Creates a Quiz

Live Demo: Devin's Codebase Wiki and Search Capabilities

Devin's Automation with Linear for Task Management

Understanding Devin's Strengths: Best Use Cases

Debates in Designing Devin: Niche vs. General Product

Building Stickiness and Defensibility in AI Products

The Technology Enabling Devin's Capabilities

AI as the Biggest Technology Shift: Hardware vs. Software Distribution

Adopting Devin in Companies: Cultural Shifts and Strategies

Startup Wisdom: Speed, Hiring, and Future Vision

Imitation Learning

This AI paradigm involves training models by having them read vast amounts of text from the internet, then teaching them to generate responses that mimic human conversation, as exemplified by early ChatGPT models.

High Compute RL (Reinforcement Learning)

A newer AI paradigm where models learn by performing tasks, receiving evaluations on whether their actions were correct or incorrect, and using this feedback to improve their decision-making and task execution over time.

Jagged Intelligence

This term describes AI's non-uniform intelligence, meaning it can be significantly better than humans at certain tasks (e.g., processing large codebases, retrieval) while being much worse at others, making its capabilities uneven.

Bricklayer to Architect Shift

This concept describes the evolving role of software engineers, moving from primarily implementing code and fixing bugs (bricklayer) to focusing on high-level direction, architecture, and problem definition (architect), with AI handling more of the implementation.

Jevon's Paradox

An economic principle applied to software engineering, suggesting that as programming becomes easier and more efficient, the total amount of software built and the number of programmers will increase, rather than decrease.

AI Agents

Autonomous systems that can make decisions, interact with the real world, take in feedback, and iterate through multiple steps to solve problems, representing a shift from simpler text-to-text models to more capable, decision-making entities.

?
What is Devin, the AI software engineer?

Devin is a fully autonomous AI software engineer designed to work on tasks end-to-end, integrating with tools like Slack, Linear, and GitHub to make pull requests and act as a junior engineer on a team.

?
How has Devin's capability evolved since its launch?

Devin has evolved from being comparable to a high school CS student to a college intern, and now to a junior engineer, though its intelligence is 'jagged,' excelling in some areas more than humans and vice-versa.

?
How does Cognition's engineering team use Devin internally?

Cognition's 15-person engineering team uses multiple Devins, with most engineers working with up to five Devins simultaneously, allowing them to hand off tasks asynchronously and focus on high-level direction.

?
How will AI impact the future of software engineering jobs?

AI tools like Devin are expected to lead to more programmers and engineers, as increased efficiency (Jevon's Paradox) will enable the creation of significantly more software and products, shifting engineers' roles from 'bricklayers' to 'architects.'

?
What skills will be more important for engineers in the AI era?

Engineers will need to focus more on understanding complex systems, product thinking, architecture, defining problems, and making key trade-offs, rather than routine implementation and boilerplate coding.

?
How does Devin help new engineers onboard to a codebase?

Devin can help new engineers by providing its internal wiki and codebase representation, allowing them to ask questions and understand how things are set up without feeling awkward about asking 'dumb questions.'

?
What kind of tasks is Devin best suited for?

Devin is best suited for well-defined tasks that are easy to verify and test, such as front-end feature requests, bug fixes, adding testing, or documentation.

?
How can companies successfully adopt AI engineers like Devin?

Companies can adopt Devin by having early adopters invest in setting it up, teaching it about their repos and processes, and letting it build a foothold by completing initial tasks, which then paves the way for wider team adoption as others see its value.

?
How can AI products build 'moats' or defensibility?

Instead of traditional moats, AI products focus on 'stickiness' by learning and building a representation of a team's codebase, stack, and process over time, and by enabling multiplayer aspects where agents grow in value through team interaction.

1. Embrace AI for Engineering

Engineers must actively use and integrate AI tools like Devin into their workflow to stay competitive and multiply their output, as AI represents the biggest technology shift of our lives.

2. Shift to Architect Role

Engineers should focus on high-level tasks like defining problems, architecting solutions, and making key decisions, as AI agents will increasingly handle implementation, debugging, and boilerplate coding.

3. Treat AI as Junior Engineer

Hand off well-defined tasks (not abstract problems) to AI agents like Devin, treating them as junior engineers you teach and learn with over time, providing feedback and steering their plans as needed.

4. Work with Multiple AI Agents

Leverage a ’team’ of AI agents (e.g., up to five per engineer) to execute tasks asynchronously, enabling parallel work and faster building, only jumping in when your expertise is truly required.

5. Continuously Learn Coding Fundamentals

Continue to learn computer science fundamentals, as understanding abstractions, logical problem-solving, and how computers work remains crucial for effectively instructing powerful AI systems.

6. Integrate AI into Workflows

Connect AI agents with existing engineering tools and platforms like Slack, GitHub, and Linear to automate task handoffs, issue resolution, and pull request generation seamlessly.

7. Start AI with Small Tasks

When adopting AI agents, begin by giving them small, easy-to-verify tasks to help them learn your codebase and build confidence before assigning larger, more complex projects.

8. Leverage AI for Onboarding

Use AI agents and their internal ‘wiki’ of codebase understanding to help new engineers onboard quickly, allowing them to ask questions and learn architecture details without awkwardness.

9. Stay Updated on AI Technology

Actively stay informed about new AI technologies and capabilities, as the field is evolving exponentially fast, and not using AI means falling behind.

10. Foster Early AI Adoption

Encourage a few enthusiastic team members to be early adopters of AI tools, as their success in setting up and teaching the AI will naturally drive broader team adoption.

11. Focus on Core Startup Principles

For founders, relentlessly prioritize moving fast, hiring exceptional talent (fighting to get them), building products people truly want, staying close to customers, and anticipating future trends.

12. Detach Personal Worth from Outcomes

As a founder, commit fully to your work and put everything into it, but avoid tying your personal emotion or self-worth to startup success or failure, fostering resilience and greater effectiveness.

13. Reimagine Processes from Scratch

Regularly question and redesign existing processes and workflows from the ground up, especially in fast-changing environments like AI development, to optimize for new capabilities.

14. Utilize AI for Research

Leverage AI to research topics, process large amounts of information, and generate content, such as creating quizzes or summaries based on online data.

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

Scott Wu

Programming is only going to become more and more important as AI gets more powerful.

Scott Wu

Devin will always be enthusiastic. We'll always be ready to put in the hours.

Scott Wu

One of the ways that we've kind of thought about Devin and building Devin is really allowing engineers to go from brick layer to architect, so to speak.

Scott Wu

Devin is best when it is working on tasks that are well defined. You want to be giving Devin tasks, not problems.

Scott Wu

I think it's often less about moats and more about stickiness.

Scott Wu

I think in terms of base intelligence, we're honestly basically already there.

Scott Wu

You should fight to, to all ends basically, you know, to, to, to, to get the folks that you really want to, to, to bring in.

Scott Wu
15
Cognition's engineering team size Total number of engineers on the team.
25%
Devin's current contribution to Cognition's pull requests Approximately a quarter of all pull requests are committed by Devins.
50%
Expected Devin contribution to Cognition's pull requests by year-end Expected to be more than half by the end of the year.
up to five
Number of Devins an engineer at Cognition typically works with Most engineers on the team work with this many Devins simultaneously.
several hundred
Devin's monthly pull requests into Cognition's production Merged into the Devin codebases every month.
10%
Average software engineer time spent on high-level directing The remaining 90% is often spent on implementation, debugging, and other routine tasks.
eight
Number of pivots Cognition went through Pivots within the domain of coding agents over the last year and a half.
18
Cognition team members who have previously started their own company Out of a total team of 26 or 27 people.