AI coding agents have been doubling the amount of uninterrupted work they can complete about every 70 days.
2
Devin's useful tasks have progressed from repetitive migrations to isolated fixes, broad repository changes, and autonomous backlog work.
3
As tasks become harder, agents need codebase understanding, human collaboration, self-testing, and the ability to ask for help when they lack confidence.
Summary
Scott Wu describes software engineering agents as moving through capability tiers every few months. He measures progress by the amount of work an agent can complete before a person has to intervene, and says coding agents have been improving faster than general AI agents, with a roughly 70-day doubling time. Devin first found adoption with repetitive migrations, then moved to isolated bugs and features, broader cross-file changes, and work assigned through Slack. More difficult tasks required repository snapshots, codebase memory, language-server information, search, codebase explanations, and better planning. Devin 2.0 adds a more iterative workflow in which engineers explore a repository with the agent, steer part of the work, and let it operate independently for the rest. Wu says the latest stage is handing Devin a backlog and expecting it to scope, test, debug, and complete tasks across repositories. He expects another large increase in coding-agent capability over the following year.
Coding agents are improving faster than general AI agents
Wu measures an AI's capability by how much work it can do without interruption before a human must intervene or steer it. He says this length of task has doubled about every seven months in general AI, while code has improved faster, at about every 70 days. That means four to six doublings in a year, or roughly 16x to 64x more coding work over that period. He contrasts GPT-3's short outputs with current agents that can handle software tasks taking humans hours.
The first dependable use case was repetitive code migration
Around the previous year, Devin's broadest adoption came from migrations such as JavaScript to TypeScript, Angular upgrades, and Java-version changes. These jobs often involve thousands of files and a known sequence of steps, but they are difficult to solve with a fully deterministic program. Wu says this was work agents could perform consistently, and it also removed an especially tedious task from humans. Cognition built playbooks so users could give Devin a clear procedure and have it follow the steps.
Memory and repository setup made repeated work practical
Repeated migrations exposed two practical needs. Devin had to preserve lessons from earlier attempts, such as remembering a special action whenever it encountered a particular situation. Cognition built knowledge or memory systems to apply those lessons to future tasks. The team also created repository snapshots that Devin could load, roll back, and use inside a clean remote virtual machine. That environment could run the repository's CI and linter before the agent returned its work.
Devin moved from migrations to isolated bugs and features
As the agents improved, users could describe a contained bug or feature instead of spelling out every step. Wu gives the example of changing a repository select dropdown so currently selected items appear at the top. These tasks usually affected one or two files, but Devin still had to set up the repository and run basic checks such as lint and CI. Wu compares this stage to having a junior engineering buddy who can take a clearly described request and implement it.
Broader changes required a model of the whole codebase
By the fall, Devin was handling broader bugs and requests that crossed several files and could involve hundreds of lines. The agent had to diagnose the problem, understand how files related, and keep its changes consistent. Wu says text alone was insufficient. Devin needed call hierarchies, a language server, git commit history, linters, and the ability to reference code across files. Slack became part of the workflow because engineers could tag Devin on an issue and ask it to investigate or build the requested change.
Complex work starts with exploration and human steering
For architecture improvements, slow functions, and unclear error handling, engineers often do not know the complete solution when they create the task. A short prompt cannot reliably produce the right result. Cognition released DeepWiki and search so people could ask questions and understand a repository before assigning the implementation. Wu describes an iterative workflow: explore the codebase with the agent, work out what the task requires, then let Devin execute it. Devin 2.0's IDE experience supports close monitoring for part of a task while the agent works independently for the rest.
Autonomous backlog work depends on confidence and self-testing
The latest stage is handing Devin many backlog tasks at once. To do that, it must identify the right repository and code area, understand the request, work across files, and decide when to ask a human for approval or clarification. Wu says confidence is part of the problem: Devin should act when it understands the task and ask for help when it does not. It also needs an iterative testing loop, with local execution, appropriate tests, debugging, and feedback on its own changes before delivering complete pull requests.
Each capability doubling changes the product interface
Wu says the apparent 2x improvements hide different engineering problems at each stage. Line completion was mainly a text-prediction task. Later versions had to work with humans in Slack, Linear, or an IDE, take feedback, plan repository changes, run shell commands, test outputs, and make longer-term decisions. The interface and the bottleneck therefore change every few months. He expects coding agents to improve by another 16x to 64x over the next year, based on the pace he describes.
"You have to be able to say, okay, I'm quite sure that this is the task and I'm going to go execute it now versus I don't understand what's going on. Human, please give me help."13:49
Who should watch
You are deciding where a coding agent can fit into a real engineering workflow and want examples beyond autocomplete.
Your team has large migrations, repository backlogs, or recurring maintenance tasks that are hard to automate with fixed scripts.
You are building an agent and need a view of the capabilities required for planning, repository context, human escalation, testing, and debugging.
