We're building trillion-parameter models and jamming context into them with string concatenation. I've been frustrated that "context engineering" is everywhere but nobody specifies what it actually means. Frameworks define the tool loop but leave injection points undefined. This is a proposal for what context engineering infrastructure should look like: - Renderable context components (tools serving UIs vs models differently) - Queryable conversations (event stream with materialized views) - Reactive injection (rules that fire based on conversation state) - Injection queue (priority, batching, deduplication) - Hookable architecture Blog post: https://michaellivs.com/blog/context-engineering-open-call Repo (early design phase, looking for collaborators): https://github.com/Michaelliv/context-engine Interested in feedback on this direction, especially: How should rules be expressed - DSL or code? What's the right integration surface for existing frameworks? |