Had an awesome time today at #apidays and met a lot of new faces and some familiar ones as well. The slides for our talk about software supply chain security are here speakerdeck.com/dasiths/buil… @NotaryProject @OpenPolicyAgent @orasproject @AquaTrivy @GrypeProject

Just launched PromptyDumpty - a universal package manager for AI agent artifacts! 📦 Install & share prompts across #coding #agents 🔄 One package format, works everywhere 🎯 Auto-detects your #AI agent 🧹 Clean install/uninstall tracking Learn more 👉 dumpty.dev

Security vulnerability assessment (VEX) using #GitHub #Copilot and #Trivy #MCP server. github.com/dasiths/ai_genera… 🔍 Runs Trivy security scans 🧠 Analyzes your actual code for exploitability 📊 Generates comprehensive VEX reports 💡 Provides clear remediation guidance

I’ve been exploring hypervelocity engineering workflows with AI agents like #GitHub #Copilot, and one fundamental challenge continues to surface: maintaining shared context alignment between developers and AI. How do we solve it? dasith.me/2025/04/02/vibe-co…

I've been doing some vibe coding experiments recently and came up with this pattern I call "breadcrumbs" that's giving me consistent results. The Breadcrumb Protocol is designed to help you vibe code by creating a shared scratchpad. youtu.be/etYG-6-9Mlk

#dapr-agents let developers design, test, and deploy agents that integrate seamlessly as collaborative services in larger systems, without reinventing microservices. I've done some experiments here. github.com/dasiths/llm-plan-…

This is a little experiment with #Autogen I did a while back to see how well it can be integrated with #MCP servers. I used the SelectorGroupChat pattern with each MCP server represented by an expert agent. github.com/dasiths/llm-plan-…

I'm really excited for the idea behind Anthropic's Model Context Protocol (#MCP) but have some reservations about the execution. There are some limitations that make it a non starter for enterprise right now. One of those is Authn/Authz. github.com/modelcontextproto…

My team have been working with #PromptFlow recently and we did some benchmarking of pf-serve to find throughput problems. I wrote up a blog post with our findings here. devblogs.microsoft.com/ise/p…

Ah @stephentoub’s annual .NET performance post is here! devblogs.microsoft.com/dotne… #dotnet

My team recently concluded a project that used #promptflow. There were some learning around optimizing for throughput and latency. We developed an awesome little throughput testing kit for pf-serve and contributed it to the Promptflow OSS repo. github.com/microsoft/promptf…

Highly requested docker file feature coming in the next version of aspire github.com/dotnet/aspire/pul… #dotnet #aspire

I wrote about a simple pattern you can follow when developing #LLM apps to prevent a certain class #prompt injection attacks with regards to user enumeration when using tools. Examples using #langchain. dasith.me/2024/05/03/llm-pro…

As long as AI systems are trained to reproduce human-generated data (e.g. text) and have no search/planning/reasoning capability, performance will saturate below or around human level. Furthermore, the amount of trials needed to reach that level will be far larger than the amount of trials needed to train humans. LLMs are trained with 200,000 years worth of reading material and are still pretty dumb. Their usefulness resides in their vast accumulated knowledge and language fluency. But they are still pretty dumb.

Legit question. What is the value of frameworks like LangChain, Autogen, crewAI, ... that basically build the same abstractions on top of a programming language that the underlying programming language already supports. For example chaining is sequential composition, agents in multi-agents frameworks are just objects (or if you want to be fancy actors) and interaction patterns are just control flow. What's wrong with just writing code. Since these frameworks are so popular, there must be some deep attraction or advantage to using them. Maybe if you would actually design a completely new language that supports these notions, that might be something. But even then I seriously doubt that the switching costs are worth it. What am I missing?

Wait… what!? Fault tolerant 2PC that is simple and commits in 1RTT? 🤯 If you missed @ChrisJe34211511 fantastic talk in #eurosys24 (#papoc24) definitely check the paper bit.ly/U2PCp

Most application developers likely dont understand ACID transactions well, but I wonder how many believe transactions are important for preventing bugs anyway. What do you think? blog.acolyer.org/2015/09/08/…

Have been experimenting with "plan and execute" agents for extending capabilities of LLM & GPT4 to consume "live" data sources. Making some awesome progress and cool demos. Check it out. I call it a "knowledge mesh" #llm #langchain #gpt4 github.com/dasiths/llm-plan-…