Discover real AI creators shaping the future. Track their latest blogs, X posts, YouTube videos, WeChat Official Account posts, and GitHub commits — all in one place.
RT Viv Ashpreet is great, this is a good faith discussion on how ai ppl define agents and how software abstractions affect builders thinking about agent systems at different times of the dev life cycle i agree with a bunch of this, have blogged a ton on “harness engineering is systems engineering” to me all of this is a meta-discussion on what level of abstraction builders should use at what time. and when they need to move between abstractions —> especially when deploying to users like infra layer always exists but not always needed to keep in my working memory if we’re designing the instruction following for the GTM agent thinking about the “harness” as a Product Creation exercise to wrap + mold model intelligence to be a great at your Task (not just coding) is a fantastic starting abstraction level for most ppl harness components here include prompts, filesystems to work with data/context, orchestration patterns via subagents, context management + compression these all matter at this abstraction level. when we’re going to prod and need to think about things like infra, then move down the abstraction chain this isn’t fully decoupled ofc, you can think about the deployment + security considerations of your design when you’re designing your “harness” ex: we and lots of other agent teams do our best to make it easy to “do the right thing” when deploying when you’re ready to go down the systems eng journey, then layer on RBAC, security policies, data durability, resumability, scaling infra, etc we ship “Production Guides” for this reason too, lots of other great teams do the same. But we rarely see a post that talks about RBAC and context engineering with the filesystem in the same mega-post. It’s fine to do it but wrong abstraction levels imo for the reader tldr: we can figure out naming the systems engineering things are important because agents are software and we should treat them as such with best practices we already know it’s good to think about...
Maybe I'm missing something, but "harness engineering" might be doing more harm than good. I've read a couple of posts on harness engineering, filesystem memory, subagent architecture. All real, all important. I've learned a lot from them. But I keep coming back to this: the
View quoted postRT john kutay Agent memory will become a first class citizen in every orgs data estate. Study how to manage it properly: Original tweet: https://x.com/JohnKutay/status/2040846889814462776
i wonder how many total eggs will go undiscovered today at these easter egg hunts
Activity on rasbt/mini-coding-agent
rasbt closed an issue in mini-coding-agent
View on GitHubActivity on rasbt/mini-coding-agent
rasbt opened a pull request in mini-coding-agent
View on GitHubRT dharmesh Re You just made my day. Late last night I watched your talk "No Vibes Allowed" at AI Engineer. It was awesome and useful. Have watched previous talks of yours as well (thanks @swyx for publishing these!). You are the real deal. In any case, when I came across this tweet of yours saying you'd watched my episode with @lennysan, dots pleasingly connected in my universe and I thought: The Internet is an amazing place. Which it is. Cheers. Original tweet: https://x.com/dharmesh/status/2040833818710487185