Making Ideas To Impact
For the last six weeks I have been running my project work through a two-agent loop, and it has changed how I think about AI assistants. Opus 4.8 plans. Gemini 3.5 executes. I sit between them as the human in the loop, and the work gets faster and cleaner than any single-agent setup I have run before.
This is what the flow looks like, what each model is actually good at, and where the loop breaks when I push it too hard.
I have been building RAG pipelines for two years. The mental model I started with was wrong, and reading Angela Shi’s article “Retrieval Is Filtering, Not Search” on Towards Data Science this week made the fix click.
The standard framing of RAG retrieval is “find the passages most similar to the query.” That framing is misleading because it imports the wrong mental model. Retrieval is not a Google-style search across unstructured text. It is a filtering problem on structured tables. The closer mental model is a SQL query, not a Google search.
This is the article that should have existed when I started. Here is what I learned, and what I am changing in my own RAG pipelines because of it.
I shipped a feature on a Tuesday that took 11 minutes end-to-end. The agent generated the happy path, ran the tests, opened the PR. I clicked merge. Done before lunch.
The same agent shipped a feature on a Friday that took me 6 more hours after the agent finished. The happy path looked identical. The difference was the last 20%.
That gap is what this post is about.
It is Monday morning. The coffee is brewing. I am writing this instead.
This is what happened last week. The five stories that grabbed me the hardest, ranked by how much noise they made across my feeds. I actually read them this time, not just the taglines.
—
Somewhere in the West, a server room is humming right now. The room is cold in a way that has nothing to do with weather. Somewhere downstream from that room, a river is running a little warmer than it was last summer. Somewhere near that river, a child is brushing her teeth. The water in her cup was clean this morning. By the end of the day, it might not be.
Last month I built a recruitment portal from scratch learn the 20% of design thinking that actually works — request form, approval flow, candidate filtering with AI, the whole nine yards. Before I wrote a single line of code, I sat through fifteen hours of interviews with HR managers, hiring managers, and candidates who had just been rejected understand requirements the way stakeholders see them.
That is the part most articles about building products skip. They jump straight to the whiteboard sketch or the workshop exercise. Those are the easy parts. The hard part is being willing to throw away your first idea after you have heard the third interview say, “that is not actually how I do my job.”
This is the story of how I actually built that portal — the conversations I had before writing code, the assumptions I had to throw out, and the things that broke in the first week. The portal handles request forms, multi-level approval, job posting, candidate registration, AI-assisted filtering, interview scheduling, psychological tests, salary offers, MCU (medical check-up), and onboarding logistics. I will show what survived and what I deleted.