Exploring the best abstraction for us to host the claude-agent-sdk. It is not straightforward. We particularly want persistent storage as I have a theory the SDK performs best with filesystem docs.claude.com/en/docs/agen...

Hints of shipmas coming around the corner!!

Gemini 3 spotted in the wild

I noticed a pattern, agent frameworks are all racing to add governance layers now, we're doing it at work

LangChain just shipped human-in-the-loop middleware + tool-call limit controls. One demo literally shows "reining in a spend-happy shopping agent"

Runaway tool calls = runaway costs

thorny decoupling issues abound. we unwittingly coupled ourselves to Svelte4 with a pnpm workspace ui package of shadcn. it's been a beast to untangle and the only solution we've really found is pinning bits-ui internally in the package and side loading bits-ui directly in our app

I haven't wrapped my head around models doing 200-300 sequential tool calls without human intervention like Kimi K2. I'm not sure I could do that so if this is SOTA we'll see a lot more of this 🧵

been using this a lot on my poorly written skills to great effect, thank you @scottspence.dev github.com/spences10/cl...

`Use pnpx claude-skills-cli to create a skill for my API client code

We're starting to test the @anthropic.com Agent SDK as THE runtime

Debugging taxes kill iteration velocity.

Fast failure only works if learning is faster.

engineeredintelligence.substack.com/p/building-s...

Is it time to go whole hog on Claude skills and drop mcp? Feels too soon

I posted this a little while back and it looks like Claire Vo just shared a rockstar masterclass on the same thing.

engineeredintelligence.substack.com/p/ephemeral-...

Vibe Analysis -> www.youtube.com/watch?v=KOr-...

Claude skills have changed nearly everything for me

Hello world