-
A copilot demo that works 90% of the time is not 90% of a product. It’s 10% of one. The remaining 10% determines whether anyone uses it.
-
Every other AI project we’re called into starts with ‘we’re building a RAG system.’ That’s never the goal. Treating RAG as the strategy guarantees expensive surprises.
-
Most agentic-coding sessions fail in the first prompt. We’ve found a middle path: spend 12 minutes on the brief, then let the agent run. Here’s the structure.
-
An agent that ships features in a clean Next.js app will fall over in a 12-year-old PHP estate. The reason isn’t model intelligence. It’s context — and legacy codebases hide their rules where agents can’t see them.
-
Three agentic coding tools dominate UK developer workflows in mid-2026. They look similar from the outside. They are optimised for very different jobs. Picking the wrong one costs a fortnight.
-
Reading every line of AI-generated code feels responsible. It also stops agentic coding from compounding. Review the plan, the inputs, and the outputs — not the diff.
