Hi HN! I built BEAR as an experiment (which might actually be useful) around a problem I’ve been thinking about recently.
When agents can generate large amounts of code very quickly, many important changes become structural rather than purely implementation. New dependencies appear, boundaries widen, or components gain access to resources they previously couldn’t reach.
These are architectural changes, but they often get lost in normal diffs, especially when an agent can introduce many edits in the same PR.
BEAR explores one way to make those changes explicit. The idea is that an agent first describes the system as a set of architectural blocks in a small IR, and then works within those boundaries while evolving the implementation.
From that representation, BEAR generates deterministic checks, and CI can surface when a pull request expands the authority of a component instead of letting that change hide inside ordinary code edits.
Still very much an experiment, but curious what people think.
When agents can generate large amounts of code very quickly, many important changes become structural rather than purely implementation. New dependencies appear, boundaries widen, or components gain access to resources they previously couldn’t reach.
These are architectural changes, but they often get lost in normal diffs, especially when an agent can introduce many edits in the same PR.
BEAR explores one way to make those changes explicit. The idea is that an agent first describes the system as a set of architectural blocks in a small IR, and then works within those boundaries while evolving the implementation.
From that representation, BEAR generates deterministic checks, and CI can surface when a pull request expands the authority of a component instead of letting that change hide inside ordinary code edits.
Still very much an experiment, but curious what people think.