Post

It's funny how my AI chats are all just "address the comments" or "read the spec and build this". Coding harnesses are all chat-centric but they definitely don't need to be.

Ray Myers

@raymyers.bsky.social

17 Aug

Yeah the Ralph pattern was smart to move to headless first, focusing on task + process rather than chat based.

Ryanne Dolan

@ry.codes

Copilot has autopilot, which forces the model to create tasks and keeps nudging until the tasks are complete. What's missing is a bunch of event sources -- new review arrived, new PR posted, etc, so that I don't need to keep nudging the model to reinspect the state of the world.

17/8/26, 12:45 pm0 reposts0 likes

Ray Myers

@raymyers.bsky.social

17 Aug

Yeah the Ralph md had part of the puzzle by being a control loop, “heres how you should inspect the state of the world before continuing”. But to get the experience you reasonably want, you’re right you need event integration, lots of glue. A term floated lately is Agent Control Plane. I’m iffy.

Ryanne Dolan

@ry.codes

17 Aug

We have such things at work, but they are completely autonomous. Coding agents are more personal, but too chatty. Wondering about a middle ground that just does things that you can personally approve or ignore.

Ray Myers

@raymyers.bsky.social

17 Aug

Haven’t tried this but people like it and there’s prompts inside to repurpose. ponytail - preferring simple solutions and minimal interaction. github.com/DietrichGebe...

github.comGitHub - DietrichGebert/ponytail: Makes your AI agent think like the laziest senior dev in the room. The best code is the code you never wrote.Makes your AI agent think like the laziest senior dev in the room. The best code is the code you never wrote. - DietrichGebert/ponytail

Ray Myers

@raymyers.bsky.social

17 Aug

I have a sketch called Prereq, which is maybe more the direction you are thinking of - basically making first class what is the layer of decisions we need to manage. Seems more optimized for big picture control.