dondon🪵

Author: com

  • fable 5. It’s cool, def thinks a lot / token burn

    fable 5. It’s cool, def thinks a lot / token burn

    I was most excited about getting features like /workflow /goal and essentially running “deep complex work” and of course being on the loop train. But there is something to be said for knowing how a model behaves and what the output you can expect. Don’t get me wrong, Fable 5 has been pushing out some work for me, but I’m not totally sold on it being so awesome / mind-blowing crazy.

    of course I ran thru my weekly usage in a couple days when it came back to Claude after being grounded by the US government… and I couldn’t help spinning up sessions to work on projects I had stalled on to see how it would go…

    no real point to be made here – it’s done some good stuff for me, but I think also because I’ve been improving my harness and implementing / riding the whole loop train đźš‚ which has been more of a step change than the model itself.

    Will share my “grovv-stack” setup soon and how I’ve been loopin’

    đź‘‹đźš‚

    Døn

  • harness eh? when’s the lasso?

    harness eh? when’s the lasso?

    in the ongoing ai development world, there is now the “harness.” all the things for your agent to use automatically so you don’t have to tell it. its mostly markdown content, but then you get fancy with scripts that run in loops and “remember” for you.

    goodtimes.

    always wondering what the next word to define an ai thing. prompt, skills, harness … maybe we are building up to horsey on the ranch.

    đź‘‹

    Døn

  • cognitive load when agentic engineering, and looks like full-stack is back

    cognitive load when agentic engineering, and looks like full-stack is back

    many other folks have written about it, but just adding my own notes. with cloudflare recently announcing a layoff of ~20% of their workforce and notably a concept/practice of moving to a “team of one” approach where engineers essentially become “full stack” … yeh bringing that back it seems. essentially loading on various types of work into one person that would have previously been many, of course with the heavy-help of Ai.

    now, full-stack wasn’t so bad, but you could definitely tell when a person leaned a specific way in their knowledge; e.g. design, devops, coding, etc.. even project management/organizational skills. now, we are seemingly standardizing some roles with Ai deep in the mix. So now now, from my perspective it’s the solutions/thinkers-heads that can capitalize and really drive value because they can “get it” by digging in and relying on their skills for comprehension, and just doing.

    Of course those with varying levels of risk aversion will go different distances, and almost like “we” are encouraging taking risks again… the ol “just ship it”

    back to cognitive load. of course the full-stackers can “do all the things” but at what cost? a big or little bug? stress? burnout? we’ll see, but probably a mix of all of it.

    feeling like I’ll have followups to this topic…

    đź‘‹

    Døn

  • recent annoying ai things: Claude Opus 4.7 – API Error: Stream idle timeout – partial response response received

    recent annoying ai things: Claude Opus 4.7 – API Error: Stream idle timeout – partial response response received

    Since Opus 4.7, I’ve increasingly ran into a stream idle timeout while using claude code via desktop or mobile apps. as of this posting, things are better, but was a rough couple weeks with roll-out of opus 4.7.

    I even got involved at https://claudestatus.com and saw similar issues pop-up.

    Seemingly was only to do with cloud containers/environments when using the apps, as using Claude Code terminal and VS Code extension worked well.

    just a little gripe post, but it gets better over time.

    đź‘‹

    døn

  • claude agent teams – agentic fun

    claude agent teams – agentic fun

    a number of ago after claude released agent teams, I dove in to check and see how the ol agent teams would work. good fun… parallel frenchmen doing their thing based off a shared to do list.

    testing with my “grovv stack” scaffolding for projects and was able to get six running around doing stuff.

    the claude cli has some cool outputs to keep you updated

    I now have agent teams setup in my settings.json for my gro\/\/ stack for all projects and it generally works well, including using the claude desktop and mobile apps.

  • always be agentic

    always be agentic

    claude code, codex, whatever, just always be running agents.

  • creating images with a six year old – monster truck edition

    creating images with a six year old – monster truck edition

    we haven’t done this in awhile, so always fun to get some crazy monster trucks generated.

    đź‘‹

    Døn

  • google api and mcp

    google api and mcp

    using google products? check the latest – https://developers.googleblog.com/introducing-the-developer-knowledge-api-and-mcp-server/

    đź‘‹

    Døn