Weekly Roundup #27 — AI Agents, RAG, and What Worked

This week was all about getting hands-on with AI agents and actually measuring what works. I published 9 posts across the blog, and the common thread running through all of them was this: the gap between what AI can do and what it reliably does is still where the real work lives.

Let me break down the numbers first. This week I covered AI agent workflows, RAG architecture, coding with AI assistants, and a few leadership pieces. The Full-Stack AI Builder category got the bulk of the content, with a couple of posts in the general AI category. Traffic-wise, the top performers were the pieces that tackled specific, practical problems rather than broad overviews.

The standout post this week was “RAG Retrieval Is Filtering, Not Search.” It pulled 11 views with an average engagement time of over 105 seconds — the highest time-on-page of anything I published. The insight that resonated was simple: most people treat RAG like a search engine, but it’s actually a filtering mechanism. You don’t ask RAG to find answers; you ask it to narrow down which chunks of text are worth passing to the LLM. That mental model shift changes how you design your retrieval pipeline entirely. The second-best performer was “Opus 4.8 Plans, Gemini 3.5 Executes — I Sit in Middle” with 11 views, which suggests people are hungry for practical multi-model orchestration patterns.

Another post that got good traction was “AI Wrote 80% in 10 Minutes. The Last 20% Took 6 Hours.” This one hit close to home for anyone who’s tried to ship production code with AI assistance. The 80/20 rule is real, and the last 20% is all the stuff AI can’t do yet: debugging edge cases, handling authentication flows, writing tests that actually cover real scenarios. I think this resonated because it’s honest about the current state of AI coding tools without being dismissive of their value.

On the analytics side, organic search is still quiet — the site is building authority slowly. Most traffic came from direct visits and social referrals, which tells me the audience is engaged but discovery through search engines needs more time. The GSC data shows impressions starting to trickle in on a few posts, particularly around AI infrastructure and embedded systems topics.

What’s next for this week: I’m digging deeper into multi-agent orchestration patterns. The Opus 4.8 / Gemini 3.5 split workflow worked well enough that I want to formalize it into a repeatable template. I’m also planning a practical guide on setting up a RAG pipeline that actually works in production — not just a demo with three PDFs. And I’ve been experimenting with a new approach to AI code review that I’ll share once I have enough data to know it’s not just placebo.

Related: RAG Retrieval Is Filtering, Not Search.

Related: AI Wrote 80% in 10 Minutes. The Last 20% Took 6 Hours.

Thanks for reading. See you next week.