long-horizon software engineering

Long-horizon software engineering refers to development work that isn't wrapped up in one question and one answer. It requires a process of planning, editing multiple parts, and checking the results. A simple errand versus moving houses makes the difference clear. A quick errand ends once you buy one item, but moving requires packing things up in stages, transporting them, and reorganizing everything in the new place in order. Coding tasks work the same way: fixing one short function is a very different level of difficulty and time commitment than building an entire new feature that spans multiple files.

For an AI to take on this kind of work, it has to look through the whole repository, plan out which parts need changes, actually edit the code, and then verify that everything works correctly. The reason recent coding agents emphasize this ability is that going beyond simple autocomplete for short code snippets, and instead being able to take over multi-day tasks that developers used to handle themselves, is what actually determines how useful these agents are in practice.

Meta introduced its coding agent Muse Code as being designed for long-horizon software engineering. This means it handles extended tasks that involve planning, editing, and verifying changes across multiple files throughout a repository, rather than just offering short code completions. The key point isn't that the answer itself is long, but that the task unfolds across multiple stages.

Try asking a coding agent something like, 'Add a complete sign-up feature to this repository from start to finish. Find and modify all the relevant files, and write tests too' — a request that can't be resolved in a single response. This lets you experience firsthand what long-horizon task handling actually feels like.