A framework gives you a callable agent. The harness is everything that happens next. After building one on top of Microsoft Agent Framework, six pieces stand out — notes on each, and what they contribute.
#MicrosoftAgentFramework#Claude#Harness
https://t.co/4aTua9xRP5
If the working method is AI-assisted, the learning method should be too. Notes from rebuilding Claude Code on Microsoft Agent Framework.
#Claude#MicrosoftAgentFramework#Harness
https://t.co/he6pmaMBVP
I keep rebuilding the same thing: a live HLS stream with the UTC clock on every frame. Player latency, recorder cuts, EPG flips — all obvious when you can just read the timestamp.
Built it from scratch twice. Third time I made a repo.
https://t.co/UryqXeHWFH
Caching a geo-restricted app sounds messy. 200 country codes, 200 cache entries, hit rate falls off a cliff.
The trick: cache by *what they see*, not *where they are*. Hash the set of regions a country resolves to. ~12 buckets instead of 200. Same hit rate, same correctness.
https://t.co/11e7zD9ZSW
Today is Children's Day in Turkey. As parents caught in the AI wave, we keep hearing: "We got by somehow — but what will our kids do?"
https://t.co/dZYor72ttf
Built a one-command setup that puts Claude Code on its own server. It picks up Linear tickets, writes code, opens PRs — no laptop required. Wrote about what broke along the way.
https://t.co/H7nPFPHvDZ
@mattpocockuk I mostly code on .NET and use the ABP framework, which applies DDD principles nicely. A similar opinionated framework in your stack might speed things up.
New flow I'm testing: Linear → Local Cyrus → Claude Code → GitHub.
Still figuring out where reviews fit vs just leaning on TDD.
Anyone else getting lost in the review step, or is TDD the real answer here?