Misatay - AI Agent Tool that Keeps You in the LoopMisatay is a VS Code extension for working with AI agents. When you build software with Misatay, you are involved and in charge. (NOTE: It only works with GitHub Copilot.) When Misatay says "All done!", you actually believe it! You're ready to:
Because: You've already reviewed the code! You were there from the beginning! Your new dev loop:
Jump to The Misatay Way for details! InstallationIt's a VS Code extension. So install it from the marketplace. In repos where you haven't used Misatay, you have to initialize it. Open the command palette (Cmd-Shift-P) and choose "Misatay: Install Agent". This will add an agent prompt and some skills to your repo. (I hope to avoid this step in the future, by keep these files in the extension and out of your repo.) Recommended SettingsTo get the best experience with Misatay, we recommend configuring the following VS Code setting: Auto Accept DelaySetting: This setting controls the delay (in seconds) after which changes made by the AI agent are automatically accepted. Setting this to 0 means auto-accept is disabled, which means you need to manually accept every change. We recommend setting this to a value greater than 0 to allow Misatay to work more efficiently. This gives you time to review changes while they're being made, but automatically accepts them after the delay if you don't intervene. To configure this setting:
Project StatusThis is an experiment, although I do try to make Misatay useful for real work. I am interested in finding ways to use AI in software development that keep the engineer in the driver's seat. Do you have ideas? Open an issue! This brings us to the project's Manifesto: ManifestoVision
Project Goals
The Misatay WayBefore you start: You should think of Misatay as a pair-programmer that you've tasked with implementing a feature. Misatay will take a crack at all the tasks for this feature, but sooner or later you will step in and review its code. Note that, like a human engineer, Misatay commits its changes to Git as it goes. So it is best to give it a feature branch to work on. Step 1: Plan First, open the Copilot Chat view in VS Code; then select Misatay as your agent. 
Now, break the project down into tasks — with the AI's help. Tasks are stored outside of the AI's context (by default, in the repo using Beads) so they are not lost. 
Step 2: Unleash Let the AI start work. Misatay will keep track of its progress, by updating tasks and committing its changes to Git. And don't worry about losing track of its changes — Misatay knows which changes go with which tasks, and you can review them easily later. 
(You can open the Task view by choosing "Misatay: Show Task Status" from the command palette (Cmd-Shift-P).) Step 3: Understand When you are ready to review some code, tell Misatay, and it will start a review session. Misatay walks you through the changes for a particular task, opening files and highlighting lines, and giving you a chance to comment or ask questions. All at your pace! 
Step 4: Revise And, of course, you can ask for changes!
Misatay tracks all this for you. Step ???: The AI Needs Help! Very often, AI agents will get stuck trying to fix something and just spin — wasting tokens and money. Not with Misatay! Misatay tells your agents to ask for help when they need it, using a special "needs help" task status to track this. 
ComparisonPlain VS Code with CopilotThe diff-based review experience provided in VS Code's Copilot Chat is excellent — but it doesn't scale to big projects. If, for example, you tell Copilot "Go make me a Minecraft" (and you've pre-approved all the tools it needs), then let it run for an hour, you'll come back to a pile of new code that you won't want to review. Misatay, in contrast, helps you break the project into tasks. You can then tell it to go to work, and it will work on those tasks while remembering what changes belong to what tasks (using Git, of course). When you come back, you can review the completed tasks one by one, in whatever order you want. Agents GaloreThe extreme opposite of Misatay is Gastown, a very interesting experiment! With agentic development, you unleash multiple (perhaps very many!) AI sessions on a project, and use some kind of harness to try to get them to do something productive. The advantage of this is clear: When it works, you end up with the software you wanted without any person having to write a line of it. The disadvantages are also clear:
Old-Fashioned Hand-CodingMisatay is, in fact, closer to hand-coding than to Gastown. Misatay has a disadvantage that keeps it from being a clear winner over hand-coding: inference slowness. It takes minutes for any AI model to do a significant code change. What's the user supposed to do while waiting? This is where the temptation to deploy fleets of agents comes from. But imagine that it takes your AI one second to "Build the authentication module with unit tests", rather than one minute. If we had this kind of performance — if the user did not have to wait very long for the AI — then I think that the Misatay approach to AI coding would be quite competitive with hand-coding and agent fleets. Sadly, we don't have this yet. But perhaps in the near future? RequirementsRequired ExtensionsGitHub Copilot: Provides the AI agent and code generation capabilities Task BackendThe default task backend is Beads, which is bundled with the extension. No separate installation is needed. |