Devin combines a language-model interface with the tools, feedback loops, and environment of a software engineer.
2
Cognition AI uses Devin asynchronously across production work, where it can create pull requests, respond to plain-English feedback, and work through Slack and GitHub.
3
Wu expects software engineering to shift toward problem solving and architecture while implementation becomes increasingly automated, increasing the amount of software people build.
Summary
Scott Wu introduces Devin through a live demo in which it builds and revises a name-memorization game from a TSV file of conference speakers. Devin plans the work, creates a React app, reads the input data, deploys a version, and responds to plain-English requests for changes. Wu then shows Devin working in Cognition AI's production repository, starting from a machine snapshot, using the team's playbook, and opening a pull request for a search bar. He explains why coding fits agentic systems: software engineering includes debugging, documentation, commands, testing, and repeated feedback, rather than code typing alone. Cognition AI now assigns several tasks to separate Devin sessions and reviews the resulting code. Wu thinks engineers will spend more time deciding what to build and how to structure it, while Devin handles more implementation work. He is also clear that the product remains early and has major challenges around speed, consistency, integrations, access, and user experience.
Devin can turn a plain-English request into a deployed working prototype
Wu opens with a name game for the World's Fair. He gives Devin a TSV file containing speakers' names and faces, then asks it to show two random faces and make him guess the matching name. Devin creates a plan, starts a React app, reads the TSV file, builds the site, and deploys an initial version. The first result has a mistake because the names remain visible. Wu gives further instructions in plain English, asking Devin to hide the names until an answer is selected, restyle the play-again button, and add a streak counter. Devin revises the site and deploys the updated game.
Devin works through the same development tools as a human engineer
Wu defines Devin as a fully autonomous software engineer with access to the tools a human engineer would use. In the production example, Devin starts from a machine snapshot that already contains the repository and development environment. It has a playbook with repository details and can work inside the team's Git environment. Devin creates a pull request, receives feedback from Bryce in plain English, adds a magnifying-glass icon using the requested library, handles a login issue, and produces a pull request that is merged. Wu also says Devin has built some of Cognition AI's API integrations, internal dashboards, and metrics tracking.
Agents add autonomous decisions to the text-completion model of software
Wu describes an earlier wave of generative AI products as text completion. ChatGPT, question-answering tools, marketing-copy systems, customer-support tools, GitHub Copilot, and Cursor usually receive a prefix and generate what comes next. Agents add autonomous decision-making, which creates a higher requirement for consistency and a new product-design problem. Devin therefore needs to manage work over time rather than only produce the next text response. Wu says the team is still learning how agent interfaces should handle parallel work, information gathering, and context.
Coding is suited to agents because the job includes many actions beyond typing
Wu argues that software engineering naturally fits an agentic workflow. Engineers inspect bugs, move through files, run commands, read documentation, start the front end, reproduce failures, edit code, and try again. When debugging a large production codebase, they use print statements, logs, monitoring, and repeated experiments to discover what happened. Each action provides feedback that guides the next decision. An agent can follow this loop because it can gather information, run code, inspect results, and change its approach instead of only completing a code fragment.
The product includes infrastructure for sessions, collaboration, and verification
Wu lists product features that do not have direct equivalents in ordinary software interfaces. Devin can use a shell, edit code, and browse the web, but it can also fork and roll back sessions, integrate with Slack and GitHub, use playbooks, work from machine snapshots, keep track of secrets, and use tools for verification. He demonstrates a newer feature that lets a developer open a VS Code Live Share session on Devin's machine. The developer can inspect the work, make an edit, and discuss the change with Devin. Wu expects the product experience to keep changing as the team learns how people should work with agents.
Cognition AI uses multiple Devin sessions to make engineering work asynchronous
Wu says Cognition AI now handles tasks in a more asynchronous way. An engineer with four tasks can give each task to a separate Devin session and let them run in parallel. He compares the sessions to enthusiastic interns: they work hard, make mistakes, and ask questions, so engineers still review and direct them. The team can also send a request from Slack, such as changing a color, then wait for Devin to create a pull request and merge it later. Wu says this works even when team members are in the gym or a car and do not have their computers with them.
Devin is intended to increase the time engineers spend deciding what to build
Wu divides software engineering into two jobs. One is deciding what solution to build, including the architecture, flows, details, and edge cases. The other is implementing that solution through debugging, functions, tests, migrations, and related work. He estimates that the average engineer currently spends 10 or 20 percent of the time on the first part and 80 or 90 percent on implementation. His view is that Devin will take on more implementation, freeing engineers to spend more time on problem solving. He compares the future role to a mix of technical architect and product manager.
Wu expects more software engineers even as implementation becomes easier
Wu does not expect automation to reduce the total amount of software engineering. He compares the change to earlier shifts from punch cards to assembly and then to C. Fewer people use the older methods, but there are more programmers overall. He argues that demand for software is already greater than what people currently build. If each engineer becomes five or ten times more effective, people may produce much more software and make software development accessible to more people. For junior engineers, he expects training to change while fundamentals such as understanding problems, knowing architectures, and handling edge cases continue to matter.
The path to Devin's broader vision still depends on improvements across the stack
In the questions, Wu describes the remaining work as broad rather than tied to one missing feature. He names speed, consistency, access, integrations, and product experience as challenges. He expects progress from several parts of the stack, including new hardware, foundation models, and agent research. Cognition AI will work on its part, but Wu says improvements across the field will also make agent workflows better. He gives no simple timetable or single solution, and says the technology is extremely early.
"Devin is a fully autonomous software engineer and what that means is Devin has access to all the same tools that a human software engineer would have when they're building something like this."01:03
Who should watch
You are evaluating coding agents and want to see what an end-to-end workflow looks like beyond code completion.
Your team wants to assign repository tasks asynchronously while keeping pull requests and human review in the loop.
You are thinking about how engineering roles and junior-engineer training may change as implementation becomes more automated.
