Daybook, 20 July 2026

Business as usual, work as usual. Kept reading The Correspondent. And today, mostly, the product was on the grind — heads-down, designing some of the multi-agentic systems and reading through the foundations side of it.
And again, right — it's product to engineering to research, that's how it keeps going. People might want different workflows, different personas to be able to handle things, or maybe very specific agents. But to build something scalable, you need to find the repetitive patterns, make it very customizable, make it a very general one — so it can be picked up and used quickly, and so we're actually able to validate it too. That's the whole thing.
There was a paper I was referring to, from Distyl AI, on agentic systems — and it's close to what I'm actually working on. Right now the agentic system I have has three elements. One is self-discovery of context — it goes and builds that semantics and ontology itself. The other one is the system, the agents, that can improve over time — but that one's more at the POC stage, so I have to keep testing it, keep throwing hard tasks at it, and see how it improves over time: through rearranging the prompts, or curating new tools, or optimizing the harness. Though right now the harness optimization isn't actually in place — so, still something I need to get to.
Settled · 20 Jul 2026
Disagree, or know something I missed? Write to mepc3197@columbia.edu · the title goes in the subject line
Related
- Open Models, Sovereign AI, Systems, Scale and Product Research and Deployment Engineeringcourseplannedengineering
- Research Engineering Alchemycourseunder wayscience · engineering
- Complex Systems, Not Integrations: What Forward-Deployed/AI Engineering Is Missingessayscience · engineering