I help organisations put AI to work in ways that stand up to scrutiny, whether that comes from a regulator, an attacker or their own board.
I tested my no-vector-database meal planner on forty real questions. Embeddings alone lost, vectors with filters won, a graph won relationships, and there is a vector database in it now.
Three tool interface choices that did more for my agents than adding agents: a precondition, a lookup and a list.
I set up a two-week trial of a new search tool. Every check passed, and my agents weren't allowed to use it.
Every integration an agent holds has a width and a height. The joint NCSC guidance on agentic AI covers the width; this model, proposed for organisations running agents at scale, adds the height, and is written up for others to challenge.
The GUI didn't kill the CLI, it repositioned it. Conversation will do the same to the GUI, and value moves down to the data.
The advantages that feel most solid, knowledge, expertise and operational excellence, are the ones AI hollows out first.
All notes