@StevenNevins

“Deploying more capital - steady lads”

loop engineers in shambles

the number of people misquoting this chart is insane It’s equivalent human hours, ie how long it would take an expert human to complete the task. It doesn’t equal how long the agent harness ran to complete the task. That might be dramatically more or dramatically less (who knows! Because it’s not measured in this chart.) Even heard an anthropic employees cite that it was harness runtime in an AI engineer talk from last week

skills are a mess in Claude code right now for the advanced features that have been released. ie, Hooks, context forking, subagents, and distributing them via plugins. Distributing a skill via a plugin changes the context isolation behavior for context: fork or subagent features vs copying and pasting the files directly into your project Running a skill inline vs using the agent fields or context: fork option changes the hook behavior. It’s basically impossible to build on top of at the moment since things just fail silently in combinations that arent the basic use cases

I’ve had this happen a handful of times now where I’m talking with someone who is actively building with ai tools and we are comparing notes and I start talking about identifying recursive loops to have the agent iterate, refine, and self improve outputs and then I realize I need to take a step back and make sure I’m not going over their head and then ask if they’re using plan mode and then they say they didn’t know about it or how to toggle thinking mode we’re so early