sometimes you elaborate the types, and sometimes, well, they elaborate you.

Types

Formal Spec-Driven Development

The increased adoption of agentic (or dynamic AI-based decision making) software has opened new challenges for verification. Most folks are relying on integration-test-style "evals" to ensure that agents do what you expect, but a handful of us are looking at more rigorously verifying agentic correctness. The idea is to adapt ideas from formal verification (model checking, proof checking, etc), and see if they can be usefully extended to the more black-box and dynamic nature of agentic software. You can see this as a special kind of Neuro-symbolic AI, applied to verifying agentic correctness. The companies that I'm aware of that are trying to solve this problem include @colimit_io, @VerifiedbyAI, and @informalinc. I also know that @headinthebox is trying to address similar problems at his new startup. It's a small community so far, but if you know others please share them so we all can learn from each other :)

Blog version: colimit.io/blog/git-inspired… 🔀 One interesting approach that we landed on applies Git's mental model to this problem. By separating viewing files (like browsing) from managing context (like Git's staging of changes), we can get two more focused modes of interaction:

Co-limits in category theory are also called direct limits, inductive limits, and right roots.