Glossary

Your lens

My own understanding of these terms, not official definitions, so that we read the Insights with the same vocabulary.

Agent: A program that is given a goal and works out for itself how to reach it. It gets there a step at a time: decide what to do next with the help of an AI model, act, look at the result, decide again. What it may touch is set in advance.
Agentic coding: Coding work done by an agent rather than typed by hand. It takes a task, works through the codebase a step at a time — reading, editing, running what it wrote, correcting from what it sees — and hands back a change. What stays with the person is stating the task and deciding what to accept.
Agentic Engineering: Professional software development with agents across the whole lifecycle — planning, requirements, architecture, implementation, testing, review, release, operation — under the same engineering discipline as before. The obligations do not move: whoever directs the agents stays answerable for what ships, and decides what to accept.
AX (Agent Experience): What User Experience is for a person, this is for an AI agent — how well the software lets it find what it needs, act correctly, and check its own result. Every codebase and every product already has one; the only variable is how much of it was designed.
Human in the loop (HITL): A working arrangement in which the machine does the work and a person decides the steps that matter. The phrase only carries information when the line is drawn somewhere specific and the system enforces it, rather than whoever happens to be watching — otherwise every setup claims it. On this site the line is drawn at irreversibility: a person starts what cannot be taken back.
Non-determinism: The property that the same input does not reliably produce the same output. Ordinary software does; a language model composes each answer by choosing among plausible continuations, so two runs come back recognisably similar, materially different, and both right. And your request is only part of that input: the standing instructions and the whole conversation so far go with it every time, and they differ from run to run.
PRD (Product Requirements Document): A document that states what a piece of software must do for the business, and why, without prescribing how it is built. It carries scope, business rules, user-visible behaviour and acceptance criteria; architecture, library choices and implementation sequencing deliberately stay out of it. Keeping the technical detail elsewhere is what lets the requirements be read and corrected by the people who own the business rules.
SDLC (Software Development Lifecycle): The sequence of phases software passes through from idea to operation. Each phase is its own competency and ends in its own artefact, which the next phase builds on. Agents can now support every phase, while a person still drives each one.
Skill: A folder of instructions an agent loads when it applies — a named procedure, often with its own scripts and reference files. Some skills the agent picks up itself, deciding from the description they carry; others run only when a person calls them by name. Either way, the procedure is written down once and can be versioned and reviewed like code.
Vibe Coding: Building software with an AI model and judging it only by whether it does what it should. The code itself is not what counts: it is read little or not at all, and its structure, tests and security play no part in the judgement.