lotharschulz

lotharschulz

@lschu.bsky.social

https://www.lotharschulz.info/

10 followers10 following43 posts

lotharschulz
@lschu.bsky.social

"Effective harnesses for long-running agents" - Neat: use two agents (an Initialiser and a Coding agent) to solve the long running agent challenge. The "Agent failure modes and solutions" section contains great troubleshooting guidance.

lotharschulz
@lschu.bsky.social

"How GitHub’s agentic security principles make our AI agents as secure as possible" - Clear explanation of how explicitness and boundaries secure Copilot agents. Tip: you may use copilot custom instructions to implement these principles (link below).

lotharschulz
@lschu.bsky.social

"How to write a great agents . md: Lessons from over 2,500 repositories" - Updated this week; includes 5 GitHub Copilot agent proposals. These were great inspiration for the agents I set up this week including a project-specific setup agent (not covered in the blog post).

lotharschulz
@lschu.bsky.social

There’s a new kind of multi-agent development I call “Flow coding”, where you orchestrate a symphony of AI agents and just let them find their groove together. You set the high-level vision, define the boundaries, then step back and watch the magic happen.

lotharschulz
@lschu.bsky.social

It’s like conducting an orchestra where each musician is an AI agent with its own specialty – one handles the frontend, another manages the backend, a third optimizes performance, and they all jam together in perfect harmony.

lotharschulz
@lschu.bsky.social

Create a weird animals quiz. I put Claude Sonnet 4 & Google Gemini 2.5 Flash to the test creating "weird animal quizzes." 🚀 Gemini (Canvas): Faster, great look. ⏱️ Claude: Comparable results, but slower. Both excellent for fast AI prototyping! What are your experiences? #GoogleGemini #ClaudeAI

lotharschulz
@lschu.bsky.social

Anthropic published a new paper on "Constitutional Classifiers", which aims to defend against universal jailbreaks. One focus is on CBRN (chemical, biological, radiological, and nuclear harms):

lotharschulz
@lschu.bsky.social

"In particular, we’re hopeful that a system defended by Constitutional Classifiers could allow us to mitigate jailbreaking risks for models which have passed the CBRN capability threshold outlined in our Responsible Scaling Policy."

lotharschulz
@lschu.bsky.social

„Not only are neurologists unlikely to be replaced by large language models any time soon, but our findings suggest that they may soon find themselves treating new, virtual patients—artificial intelligence models presenting with cognitive impairment.“ medicalxpress.com/news/2024-12...

medicalxpress.comLeading AI chatbots show dementia-like cognitive decline in tests, raising questions about their future in medicineAlmost all leading large language models or "chatbots" show signs of mild cognitive impairment in tests widely used to spot early signs of dementia, finds a study in the Christmas issue of the BMJ.
lotharschulz
@lschu.bsky.social

Just stumbled upon Google’s new AI tool, Whisk, but sadly it’s not yet available outside of US. Seeing innovation roll out and having to wait is always a little bittersweet. Hoping it becomes accessible soon so we can see what it brings to the table! Who else is waiting for it?

You are up to date.