agentic engineering scorecard→agentic engineering scorecard→practice - custom linting rules that inject in-repo documentation as context→video - harness engineering from OpenAI - 2026-04→model attention is scarce

model attention is scarce
link not tracked is link not tracked

Backlinks

video - harness engineering from OpenAI - 2026-04

At 5:10 it is not just human attention that is scarce but also model attention is scarce

see in context

At 7:30 writing things in ways that is native to the agents and structuring them in a way that is respectful of the fact that model attention is scarce

see in context

At 11:30 auto compaction has gotten significantly better and so context people don't start new as much and so context gets paged out, so you have to figure out a way of ensuring that the context remains
...
It sounds like he's saying that we don't need Ralph loops as much, anymore, but on the flip-side he says we should be mindful that model attention is scarce.

see in context

At 13:00 he talks about adding linting rules that catch specific issues. For example, on every request, it should always have a retry and a timeout

Another static analysis check that he suggests is forcing all of the files to be less than 350 lines because model attention is scarce. Feels like the types of constraints you would add with Rubocop

see in context