AutoDev is a fully automated, AI-driven software development framework from Microsoft, designed for the autonomous planning and execution of software engineering tasks. A user defines an objective — for example, testing a specific method — and AutoDev’s AI agents pursue it by performing actions in the repository: writing a test file, running it, reading the failure log, retrieving more context, fixing the file and re-running the tests until they pass, with no developer intervention beyond setting the objective.
The paper introducing it, AutoDev: Automated AI-Driven Development, positions it against AI coding assistants integrated into IDEs such as GitHub Copilot, which it describes as mostly suggesting code in a chat interface while leaving developers to run tests and validate the output themselves. It describes AutoDev as extending AutoGen beyond conversation management so that agents act on the repository, and as LLM-agnostic, allowing models of different sizes and architectures to collaborate on a task.
Capabilities
Users first configure rules and actions in YAML files, enabling or disabling specific commands and defining the number of agents, their responsibilities and their available actions — for example a “Developer” agent and a “Reviewer” agent. The Conversation Manager initializes and maintains the conversation, parses each agent response into a command and arguments, checks it against the user’s permissions, and adds a structured summary of each result to the conversation; it ends the conversation when an agent issues stop, a user-defined iteration or token limit is reached, or a problem is detected. The Agent Scheduler decides which agent acts next using round-robin, token-based or priority-based collaboration.
The Tools Library abstracts low-level commands behind simple ones: file editing (write, edit, insert, delete, down to a range of lines), retrieval from grep, find and ls up to embedding-based retrieval of similar snippets, build and execution (build, run), testing and validation (test, syntax, linters and bug-finding tools), git operations with fine-grained permissions such as allowing only local commits, and communication commands talk, ask and stop. Every command runs inside the Docker-based Evaluation Environment, which returns standard output and errors to the conversation.
Adoption & Ecosystem
In a pilot study developers used AutoDev as a CLI command while watching the conversation in VS Code; the ask command, which lets agents request user feedback, was added at a developer’s request during that study. The authors plan to integrate AutoDev into IDEs as a chatbot experience and into CI/CD pipelines and PR review platforms, so that developers can assign tasks and issues to it and review the results in a PR system.