Working definitions

Using an LLM to produce software that the person directing it cannot meaningfully read, evaluate, or maintain. The distinguishing feature isn't how much code the model wrote; it's that plausibility and surface behavior stand in for understanding and verification.

Used precisely, vibe coding describes this process. Applied indiscriminately to any AI-assisted software, regardless of how it was understood or verified, it stops being a description and becomes a slur.

The deliberate choice to use a particular AI capability for a particular kind of work, under stated constraints. It is not a binary identity or a one-way stage of progress: people and organizations can adopt, limit, defer, reverse, or reject different uses based on access, evidence, risk, labor, or values.

A model operating within an execution loop in which it can choose actions, observe their results, and continue toward a goal with some delegated authority. An agent is not defined by a personality or chat interface; its agency is bounded by its tools, permissions, context, and stopping conditions.

The software and operating constraints through which an agent receives context, takes actions, and is evaluated. A harness can assemble context, expose tools, grant permissions, manage state, retries, and stopping conditions, and record evidence for verification. The agent performs the work; the harness makes that work repeatable, observable, and governable.

Engineering software with agents as active participants in a process the engineer understands and remains accountable for. Agents may plan, implement, review, and test, but intent, architecture, verification standards, and release decisions remain legible to and owned by a human.

Establishing, with evidence appropriate to the risk, that an output satisfies explicit claims and constraints. Seeing something run once or finding its behavior plausible is observation, not verification. Tests, inspection, source checking, reproducible exercises, and independent review can all contribute; the required evidence should not be chosen solely by the system whose output is being verified.

Work presented as finished despite being incoherent, unreviewed, or materially indifferent to the standards of its medium. In software, it appears as inconsistent patterns, violations of language or system fundamentals, and maintenance costs pushed onto the next person. Humans produced slop before generative AI; AI only made it faster and cheaper to produce at scale.

"Slop" becomes a slur when it substitutes for identifying those specific failures—especially when it dismisses a person or all AI-assisted work.