Dasith Wijesiriwardena @dasiths

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

@_philschmid This is an example of context engineering for AI agents. dasith.me/2025/04/02/vib…

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

@vasuman Yes the core principles of spec driven dev and planning apply regardless of the tool you use. There is a bit more here. dasith.me/2025/04/02/vib…

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/vib…

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-pl…

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-pl…

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/modelcontextpr…

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/promptflow…

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

@nickchapsas I wanted to reach out regarding the practice of running migration as part of your app init which is considered a bad practice (concurrency, elevated permissions). A caveat in the video description would help teams. See github.com/dotnet/efcore/…

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/prom…

@davidfowl Haven't done this for .NET 8 but would keyed services help with this scenario where you need a tenant aware IOptionsMonitorCache for example? as shown here stackoverflow.com/questions/4959…

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

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.

Pedro Domingos @pmddomingos

2 years ago

Interesting how in all these domains AI is asymptoting at roughly human performance - where's the AI zooming past us to superintelligence that Kurzweil etc. predicted/feared?

240 145 982 1.1M 376

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?

keyboard_arrow_left Previous keyboard_arrow_right Next

@reubenbond @headinthebox (cont) with strong opinions and build features like serving (http server) on top. They are great for building POC like things but don't offer the engineers enough control to leverage mature patterns or eject from the pigeonholed architecture enforced by the LLM agent frameworks.

@reubenbond @headinthebox Having worked with "big" customer recently, most teams that develop LLM solutions aren't cross functional. Most orgs still deal with DS folks who do most of their work on Jupyter notebooks. Frameworks like PromptFlow try to solve this by giving large building blocks (cont...)