I have spent ten-plus years hands-on in data and AI, and the arc matters more than the total: applied data science first, then AI engineering, and now agents together with the data infrastructure agents actually need. Each step approached the same problem from a different side.
Applied data science taught the part that is easy to skip: a model is a small component inside a much larger question about what an organisation already knows and where that knowledge is written down. AI engineering taught how quickly that question becomes load-bearing once something is in production and answers are going out unsupervised. Agent work made it the whole job — an agent is only as trustworthy as the record it reads from.
What the work is
Context management, taken literally. Where a fact enters a system. What holds it, and what that store costs to correct. What an agent is permitted to retrieve, and what it is answerable for. The interesting failures are almost never retrieval failures: they are two true-looking answers with nothing in the system recording which one is current.
The mechanisms converge across organisations — the same handful of failure shapes, in the same order. The fit to a given organisation's sources, habits and people does not converge at all, and that gap is where the work actually is.
What is here
Two kinds of piece, one stream: short technical notes derived from real work, and longer sourced guides. Every piece carries both the date the thinking happened and the date it went public, so you can tell how old an idea is rather than how recently it was posted. Every statistic carries its source and a grade, because I have been wrong about one in public and would rather you could check than take my word.
The writing is the whole of it, and/subscribe explains what does and does not arrive.