It's about the hardest agentic problem that has measurable results and will be adopted (because it's an engineering culture) that I can think of. RunWare have some good insights on this - much of the problem is context layer improvisation and repetition, and the only way out is an awful lot of diagnostic development.
And anything less than 90% accuracy on causal analysis is more work than doing everything by hand.
I am a middle aged man who lives on a sailing boat. A decade in early stage / founding without an exit that's closer to a car crash. Still trying, I don't have the temperament for consulting or corporate.
I would really like to see more people down my prompt engineering rabbit hole of trying to use ontology / compilation tricks to produce prompts from models rather than curating text. I have this working pretty well for complex tool interaction, without a JSON schema or template in sight.
Believe it or not, agents can run things like this:
e3{p13=e4(p14), p38=$} returns [e3] · Incremental collaboration signal (what changed and when). Poll to stay current; acknowledge so the server can advance your cursor · optional params: p38
And anything less than 90% accuracy on causal analysis is more work than doing everything by hand.