Hello!
Lately I've been circling the idea that turning tasks into reproducible agent outcomes is becoming a genuinely valuable skill. That's the thing worth getting good at: not just prompting an agent once, but defining work precisely enough that you get the same reliable result again and again. Why does this matter now? Because we've gone from a single agent finishing a coding task feeling like magic to orchestrating thousands of them across devices. I literally posted about "a thousand agents, in your pocket" this week... large scale agent coordination running across multiple machines, isolated workspaces, custom agents, steered by text, voice, or image. But coordination at that scale only works if the underlying work is reproducible. That's where things like agent roles come in: a coordinator breaks work into pieces, implementers write the code, verifiers check it against acceptance criteria. And intelligent handoffs let agents pass work forward carrying only what matters. None of that holds together if each run is a fresh roll of the dice. So the skill isn't babysitting one agent. It's writing specs clear enough that a task becomes an outcome you can trust, delegate, and repeat. Get that right, and a thousand agents in your pocket actually means something.
What is the tech stack of Ask LukeW feature?
The Ask Luke feature is built on a complex tech stack that has evolved considerably since its launch in April 2023.
At the core, it runs on SvelteKit with Supabase for authentication, database operations, and file storage. The whole system is deployed on Vercel at https://ask.lukew.com. For the UI, I use Svelte's component-based architecture with TailwindCSS for styling, Bits UI for UI components, and Svelte Sonner for notifications.
The real complexity comes from the AI pipeline. I don't rely on any single LLM or AI model. Instead, I use multiple AI models for different tasks - speech-to-text conversion, audio diarization, topic extraction, content summarization, PDF parsing, and more. Some models are self-hosted while others use APIs from providers like OpenAI, Anthropic's Claude, Mistral AI, Google's Gemini, Cohere, and AWS Bedrock.
The system uses a dual retrieval approach - both embedding search (for semantic similarity) and keyword search run simultaneously. Results are then ranked and diversified before being fed to the generation model. The vector-cosine-similarity library handles the semantic matching.
On the development side, I use TypeScript for type safety, Vite for the build process, ESLint and Prettier for code quality, and Sentry for error monitoring in production.
I've built an admin system that lets me quickly test new models and prompts before deploying them, since the AI space changes so rapidly and I'm constantly evaluating better models for each step of the pipeline.
If you want to dive deeper into how and why I built this system, I recommend checking out my "How AI Ate My Website" presentation in the Presentations section of my website.