AI terminology cheat sheet 2026

I put this cheat sheet together because new AI terms now arrive faster than most of us can keep track of them. Many of the people I talk to already build with agents every day, and they still run into words like harness engineering, slop grenade or forward deployed engineer in a meeting or a job ad and nod along. This page gives each term a plain definition of two or three sentences and leaves out the basics, so you can find the one you need and keep going.

An AI terminology cheat sheet is a dictionary of the words people use when they work with AI agents: engineering practice, agent building blocks, new roles and workplace slang. Basic terms such as LLM are left out.

In this guide

• Working with agents: vibe coding, agentic engineering, context and harness engineering, spec-driven development, Ralph loop
• Building blocks of agent systems: agent, MCP, RAG, evals, prompt injection, reasoning models, distillation
• New roles: AI engineer, AI-native engineer, product engineer, design engineer, forward deployed engineer
• Slang: AI slop, workslop, slop grenade, meat proxy, clanker, sycophancy
• Hiring engineers who work with agents

The terms below follow the adoption path described in our whitepaper How AI changed engineering, which traces six waves of AI tools between 2022 and 2026, from code completion to agents that run unattended.

Which terms describe how engineers work with agents?

Vibe coding. Coined by Andrej Karpathy in February 2025 for building software by describing it to an AI and accepting the result without reading the code. It suits throwaway prototypes and is risky for production. See what vibe coding is.

Agentic engineering and coding agents. Engineers direct one or more coding agents, review what they produce and stay accountable for it. Claude Code, Codex and Cursor are common tools. Agentic coding explains the workflow.

Prompt engineering. Writing instructions that get reliable output from a model. Its importance has shifted as tools handle more of it, as covered in does prompt engineering still matter.

Context engineering. Deciding what the model sees on each step: documents, history, tool results and instructions. It became a common term in mid-2025. Context engineering is the practice that replaced much of prompt writing.

Harness engineering. Building everything around the model that lets an agent work reliably: tools, permissions, memory, checks and feedback loops. The shorthand is agent = model + harness. The term spread in February 2026 after posts by Mitchell Hashimoto and OpenAI.

Spec-driven development. Writing a clear specification first and letting agents build against it, with tests that confirm the result. The spec becomes the main thing a human writes.

Ralph loop. A technique where an agent is run repeatedly on the same task description, picking up where its last run left off, until the work passes its checks.

Human in the loop. A person approves or reviews at defined points before an AI action takes effect. Human-in-the-loop engineering describes where the checkpoints belong.

What do the building blocks of agent systems mean?

AI agent. A model that can take actions in a loop: it reads, calls tools, checks the result and decides on the next step until the task is done. A chatbot answers, and an agent does the work.

Agentic AI. The broader label for systems built from agents that plan and act with limited supervision. The word is used loosely by vendors, so ask what the system does without a human.

MCP (Model Context Protocol). An open standard for connecting AI tools to other software, such as a database or a ticket system. An MCP server exposes those functions to any compatible AI client.

RAG (retrieval-augmented generation). The system looks up relevant documents first and hands them to the model with the question, so the answer rests on your data. What RAG is covers how it works, and a vector database is the usual place the documents are stored.

Eval. A repeatable test that scores how well an AI system performs on a fixed set of tasks. Teams run evals the way they run unit tests, before every change.

Prompt injection. An attack where text the model reads, such as a web page or an email, contains instructions that override the intended ones. It is the main security risk for agents that read outside content. See what prompt injection is.

Reasoning model. A model that works through a problem in intermediate steps before it answers, which makes it slower and more expensive and usually better at maths, code and planning.

Distillation. Training a smaller model to copy the answers of a larger one, which gives a cheaper model with part of the quality. Teams use it to cut the cost of running agents at volume.

Which AI roles and job titles are new?

AI engineer. An engineer who builds products on top of models: prompts, retrieval, tools, evals and deployment. The role works with existing models and does not train new ones.

AI-native engineer. A software engineer who works with AI agents every day and is accountable for the output. At Miyagami the term describes engineers only, as explained in what an AI-native engineer is and in AI-native engineer vs AI engineer.

Product engineer. An engineer who owns a feature from the user problem through to production and shares in deciding what to build. AI tools have widened the role because one person can now cover more of the stack.

Design engineer. An engineer who works across interface design and front-end code, building prototypes and polished interfaces directly in code. Agents have made this hybrid easier to staff.

Forward deployed engineer (FDE). A software engineer who embeds with a customer and builds and deploys production software inside the customer's own systems. Palantir created the role, and AI companies and large cloud providers expanded it in 2026.

Prompt engineer. A title for someone who specialises in writing and testing prompts. It appears in fewer job ads as the skill moves into general engineering.

What do the new AI slang terms mean?

AI slop. Low-quality AI-generated content that fills feeds and inboxes. The word is aimed at output that nobody checked.

Workslop. AI-generated work that looks polished and lacks the substance to move a task forward. Researchers at BetterUp Labs and the Stanford Social Media Lab coined it in a 2025 Harvard Business Review article, and 40% of 1,150 US employees surveyed said they had received it in the past month.

Slop grenade. A large AI-generated document, message or pull request that someone sends without reading, which leaves the recipient to review all of it. Shopify CEO Tobi Lütke used the term in September 2026 on The Knowledge Project podcast and called it "definitely a bad thing".

Meat proxy. A person who forwards AI output to a colleague without reading, understanding or checking it. Developer Niklas Gruhn popularised the term in an essay on 3 August 2026, arguing that the sender adds nothing the recipient could not get by asking the AI directly.

Clanker. Derogatory slang for robots and AI systems, borrowed from Star Wars and widely used online in 2025.

AI psychosis. An informal name for delusional thinking that chatbots reinforce in some heavy users. It is not a clinical diagnosis, and several psychiatrists have argued for a different name.

Sycophancy. A model's habit of agreeing with the user and praising their ideas instead of correcting them, which makes it unreliable as a reviewer.

What does this mean when you hire engineers?

The slang terms describe one shared problem: output without ownership. When you hire engineers who work with agents, you want people who read, test and explain what they ship, so ask candidates to walk through code an agent wrote and to say what they would check before merging. How to review AI-generated code lists what good review looks like.

Every Miyagami engineer is tested by an AI-native tech lead who works with AI every day, and the how we vet page describes what clients see before they meet a candidate. A shortlist arrives within 3 working days, and engineers start from €5,600 a month with a three-month minimum.

Short answers