Michael Baird

If you're a big company selling agentic hosting and you don't have a terraform provider you're doing it wrong. if you're any company publishing an agentic framework and you don't support convenient local dev features like streamlit automations, a local api server, easy local database and cache access and local eval automations -- you're doing it wrong. Honestly this is one thing Google has gotten very right with their ADK. It is indubitably the best agent framework thus far, no question.

3

It's about time for better, healthier communities for developers to engage with. Introducing Commit Confidential: Secret-Shopper style audits of developer and OSS communities that leverage evidence-based, standardized evaluation criteria that quantify the overall health of a community... from the viewpoint of the contributor.

1

Dear Claude, please create a mechanism within Claude/Cowork for org members to create plugins with skills and commands and have the ability to share them org wide without requiring an admin uploading it (no, Projects aren't the same thing). If Cowork was built in 10 days, this small capability should be what, 2 most? One shot, no bugs, please, thank you!

22

I now have my own operating system (like Windows 11 or iOS) that integrates into my own cloud. Here is an update on Vibe Coding with Claude. Overall, it's not looking great for most developers. I've moved from my cloud environment (email, doc storage, credentialing, etc) to the operating system level, an area I have little expertise. The current result is a functioning OS that is integrated into my cloud architecture for authentication that I can develop from my phone, commit the code from anywhere and have a new ISO file waiting in S3 in about 10 min. There were limitations, though. An early architectural decision proved wrong and required a full restart of the OS project. This suggests that while the model is capable of assisting with implementation, architectural guidance still needs human oversight. There were also a few instances where the model appeared overly confident while pursuing an incorrect path. In those cases, the changes had to be manually reversed. More broadly, domain knowledge remains essential. The model accelerates implementation, but understanding the underlying technologies is still necessary in order to guide it effectively. I've had to pick up quite a bit about OS drivers and greeter's this weekend. Token generation speed did not meaningfully constrain progress. The dominant time cost remains verification: ensuring that systems actually work as intended. Claims that higher token throughput materially improves outcomes are greatly overexaggerated. Finally, small implementation issues still arise and require debugging and manual correction. All in all, the technology is powerful, but does not eliminate the need for technical judgment quite yet. Next step, connecting local SLMs and cloud based LLMs into the operating system so the whole thing is AI driven. #AI #VibeCoding #Tamaton #LLM #ClaudeCode #Opus #Anthropic

38

14 Comments

"Stop using Pull Requests" Andrea Laforgia His latest Substack article on Pull Request is pure gold. Well researched, well written. I wish I would be able to write such articles... My copy-past outtake: Pull requests have become the default code review mechanism in software teams everywhere. But the pull request was invented for open source projects, where strangers contribute code to repositories maintained by people who don’t know them. => When private teams of trusted colleagues adopt the same process, they import a model designed for low trust into an environment that should operate on high trust!!! The result is a post-development inspection system that the evidence says catches few bugs, introduces enormous waiting time, incentivises large batches, and fractures teams into isolated individuals. Link to his article: https://lnkd.in/gbyNJazj

88

15 Comments

Excellent review of the history of Frontend Architecture. Like a lot of the decisions we make, there are tradeoffs. https://lnkd.in/gb6qDzBg

7

I was watching a “non technical” leader at my org vibe code a web application. Great to see Claude code reaching past my engineering teams. We started to play it forward - who manages this? Where will we host this? Ownership and bug reporting? - and you quickly realize that AI makes the building of a thing cheaper but not necessarily the running and operating of the thing. A few points: 1. AI can build software incredibly fast. But it still can’t operate it. That tells me we need AI to solve operations. And quickly. 2. If everyone is a builder, then Platform Engineering is going to need to broaden its target market. Massively. Like, the entire org is now needing infrastructure. That’s enormous but tells me that PE teams are going to be critical. Which means getting AI to take on PE roles is also key and probably an underserviced area of focus. 3. We are likely heading towards an open source model here where employees will be maintainers of tools they build. We will need to establish internal standards for promoting and professionalizing internal tools as they proliferate. Security nightmares aside. 4. Pretty much everyone in your org now needs to understand git and have a GitHub (or equivalent) account in the next 12 months if we really lean into this. Or, at the least, their personal AI assistants do. 5. Yegge called this out a while ago. This is where he thought junior engineers would go; they would become the army of technical talent that would onboard non-technical members and support internally. I’m starting to catch up to his thinking here.

54

8 Comments

Engineers that are great at debugging and finding root cause during a production outage have mastered removing 50% of the hay at a time until they find the needle. Engineers that dive straight in to find the needle in the haystack are the ones that struggle. The more complex a system is the more critical this skill becomes.

52

6 Comments