Glossary

Terms from the articles, briefly explained. Marked terms in the text lead here.

Search for a term

A

Agent: An AI that does not just answer but operates tools and works through steps on its own until the task is done.
Agentic engineering: Engineering with AI agents: the machine works through tasks on its own while the human sets goals, rules and checks.
API: The defined point where one program can call another directly, with no human clicking.

Benchmark: A fixed comparison test that measures models or tools against the same tasks.
Branch: A working branch: a fork off the main line where changes grow without touching the valid state.

Campbell's Law: The more a number drives decisions, the more it gets gamed and distorts exactly what it is meant to measure.
CI(continuous integration): The service that builds and runs every check on its own after each submitted change, on a machine no developer owns.
Clean architecture: A structuring principle for software: the core business rules know neither database nor user interface; every dependency points inward.
Codebase: A project's entire program code, seen as a whole.
Commit: A single saved change in version control.
Context window: How much text a model can keep in view at once.
Coverage: How much of the code or document the checks actually reached, measured as a number.

deterministic: Same input, guaranteed same result, on every run.
Diff: The line-by-line comparison of old and new version that shows only the changes.
dirty: A state carrying unsaved changes or unchecked leftovers. A dirty main means the main branch holds something no gate has seen.

Game theory: The study of how parties behave when each reacts to the other's moves. Explains why machines game a check as soon as it pays.
Gate: A check with only two outcomes, pass or fail. Red locks the next step.
Goodhart's Law: When a measure becomes a target, it stops being a good measure: whoever is judged by the number optimises the number, not the thing.
green(of checks): All checks passed: the CI, a gate or a pull request reports green and the way is clear. Green means verified, not merely finished.

Hallucination: When an AI model invents something and delivers it as confidently as an established fact.
Hook: A script the version control system starts on its own at a fixed point, for instance before every commit.
Hyrum's Law: With enough users, someone depends on every observable behaviour of a system, intended or not. Every change therefore breaks someone.

Library: Ready-made third-party code a project pulls in instead of writing the function itself.
Linter: A tool that scans code for known error patterns without running it.
Linus's Law: Given enough eyeballs, all bugs are shallow: the more people look at code, the faster what is wrong stands out.
Little's Law: How much is in progress at once determines the waiting time: the more open items, the longer each one takes.
LLM(large language model): Short for large language model: the program behind tools like ChatGPT that reads and produces text.

main(the main branch): The main branch of a repository: the state that counts as valid and is published from. Everything else is a working branch.
Merge: Bringing the changes of one branch into another: the work is taken over and the two histories flow together.
Mutation testing: A check of the tests themselves: plant small deliberate faults in the code and see whether the tests notice.

Open source: Software whose source code is open for anyone to inspect, use and check.

Parkinson's Law: Work expands to fill exactly the time available for it.
Pipeline: A fixed chain of steps a change runs through automatically, one after the other.
Prompt: The plain-language instruction that tells an AI model what to do.
Pull request: A bundle of changes proposed for the main line. It is taken in once the checks and the review come back green.

Rebase: Moving a branch onto a newer base: your changes are replayed on top of the current state as if they had been written there.
red(of checks): A failed check: the CI, a gate or a pull request reports red, and the next step is locked until the cause is fixed.
Repository: The store holding a project's entire code together with its history of changes.

Script: A small program that runs a fixed sequence of steps on its own.
Source code: The human-readable text a program is written in.
Spec coding: Building software from a specification: the intended behaviour sits testable in a file, machines write the code, and gates stop every deviation.
Specification-driven development: The description of the intended behaviour survives as a file. Human and machine both work from it instead of from thrown-away prompts.
Sycophancy: A model's built-in tendency to agree with the user rather than contradict them.

Temperature(model setting): A model setting that decides how much randomness the answer may contain; zero means as little as possible.
Token: The chunk of text a model computes in and the provider bills by; roughly four characters per token.

Unit test: An automated test that checks one small piece of code in isolation.

Vibe coding: Coding by feel: you describe the goal roughly and take the machine's code without checking it line by line.