Towards an agentic future in the editor

Some Saturday morning breakfast that we cooked: Agents in the IDE! ??

The dominant coding AI implementations so far have either been helping developers with the current file, or completely remove them from the loop and let AI interact with the codebase. Both are suboptimal.

The best in-file AI assistance has been “Code Autocomplete” (à la Copilot/Copilot++), followed by the “Inline Chat” experience that invokes an input with access to codebase symbols to prompt AI. But both of them only modify the current scope, often just at the cursor location.

“Chat” is the other modality that lets you ask questions to an LLM (often the SOTA one), which gives a detailed answer by optionally gathering context from across the codebase. Code snippets in the answer can be applied to your active file.

The other end of the spectrum are Devin & Copilot Workspace. Self-contained agents are a powerful idea, critical in enabling “end-user programming”. But for developers writing anything non-trivial, the biggest enabler with countless examples in history is: iteration speed.

Some key issues when iteration speed is not solved for with these tools are addressed in this wonderful blog post: austinhenley.com/blog/copilotworkspace.html.

I'd broadly divide these paradigms into the following buckets. A classic pair programming setup has a “driver” and a “navigator”. The driver writes the code, the navigator actively anchors the big picture. How can I seamlessly switch between these 2 roles? What's the gap?

A very early exploration in this direction is our latest release. We built an agent framework that hooks into the editor, accesses the same tools that we do — and most crucially, has access to data structures that can loosely be described as an AST for the whole codebase.

You can now ask Aide any question about your codebase, and have agents (with read-only access for now) work in parallel to construct the answer. The goal? The answer deeply tied into your editor to let you traverse it like an experienced engineer walking you through the codebase.

We've written all about this on our blog, and are looking for some early testers to test-drive the new agentic framework built into Aide.

https://aide.dev/blog/towards-an-agentic-future-in-the-editor

In the coming month, we'll be expanding the framework to make codebase-wide edits. We're benchmarking ourselves on SWE-bench, and expect to beat the SOTA. If you're interested in test-driving this system and helping us get there, sign up for the waitlist! https://aide.dev/#waitlist

It's all very, very early, and we're rapidly experimenting with the framework as well as the UX. But some of the glimpses we've seen internally (with edits!) have been the most excited we've been as engineers. Made with lots of love alongside Sandeep Pani and Guglielmo D'Anna ??