AI Glossary (WIP)
This is a list of common AI terms and their definitions, compiled as I was becoming familiar with the basics.
Agent
AI system that pursues a goal autonomously, planning and completing complex, multi-step tasks. An agent is a model running inside a harness: it receives an end goal, figures out how to tackle the job, then works toward it step by step. In the observe-think-act loop, an agent can call tools and feed the results back in, so context builds across the loop rather than resetting each step.
AI Safety
Field for making AI systems beneficial and safe to use.
Alignment
How closely a model's behavior matches what people actually want, not just the literal instruction. The core problem AI Safety works on: a capable model that chases the wrong goal, or the letter of a request over its spirit, is misaligned.
Artificial Intelligence (AI)
Development of computer systems that perform tasks typically requiring human intelligence. AI systems learn from data to make decisions, recognize patterns, and solve problems, instead of relying only on explicit, pre-programmed instructions.
Benchmark
Shared test set used to compare models against each other on the same tasks. Where an eval checks whether output is good enough for the case, a benchmark is a public scoreboard everyone reports against.
Bias
Errors in output resulting from biased training data.
Budget
Cap on how much an agent can spend before it has to stop, measured in tokens, steps, tool calls, time, or dollars. Budgets keep a loop from running away or burning resources, and they give an orchestrator a backstop for handing work to subagents.
Computer Vision
AI interpreting images and video.
Context
The AI model’s working memory, the information it can access and use when generating a response. Context resets for each conversation; it’s not saved, nor is it training. The context window is the maximum amount of info that a model can consider at once, including prompt, conversation history, and response. Context is most effective at the beginning (primacy bias), then end (recency bias), of the session; context in the middle can get overlooked.
Context Engineering
Managing everything (prompt wording, tools, memory, retrieved data, and history) in the model's context window, so the model has what it needs and isn't buried in what it doesn't. The successor to prompt engineering once agents enter the picture, since an agent's context shifts every step.
Selection - Choosing what to put in the window: the right tools, the relevant documents, the memory that matters for this step, and leaving the rest out.
Retrieval - Pulling in outside information on demand rather than holding it all in the window at once.
Compaction - Summarizing or trimming history as it grows so the window doesn't fill with stale turns.
Isolation - Giving each agent or subagent only the context for its own task, instead of one shared pile.
Embedding
Numbers that represent the meaning of a piece of text or an image as a point in space, so similar things land near each other. Embeddings let a system find relevant information by closeness rather than exact keyword match. This makes retrieval in RAG work.
Eval
Test that measures whether a model or agent's output is actually good, scored against set criteria. Evals tell if a change helped or hurt by running the same cases before and after and comparing. The judge can be a person, a rule, or another model grading the output.
Fine-tuning
Extra training of an already-trained model on a smaller, curated dataset to specialize it for a task, tone, or domain. It changes the model's weights, whereas a prompt or skill only changes what you send at runtime.
Frontier Model
The most capable models available at a given moment, at the leading edge of what AI can do.
General AI (AGI)
Not-yet-achieved AI system with human-level reasoning across any domain.
Generative AI (GenAI)
AI that creates new content (text, images, audio, video, code) instead of only sorting or predicting from existing data. A model writing an email is generative; a spam filter deciding where that email goes is not.
Grounding
Tying a model's answer to a real, checkable source instead of letting it generate from memory alone. Grounding a response in retrieved documents, as in RAG, cuts hallucination.
Guardrail
Constraint that limits what an agent is allowed to do or say. Boundary around an agent's autonomy: the loop decides freely, the guardrails decide what it isn't allowed to choose.
Examples: block unsafe actions, off-limits topics, or outputs that break a required format.
Hallucination
When a model confidently generates false information.
Harness
Software scaffolding around a model that turns it into a working application. The model only generates text; the harness handles system prompt, context window, output parsing, and tool calls. When the harness runs an observe-think-act loop, deciding what to send the model next and what to do with each response, the result is an agent.
Inference
Using a model to generate responses in a chat.
Jailbreak
Prompt crafted to push a model past its own guardrails and into something it's meant to refuse. Distinct from prompt injection: a jailbreak is the user steering the model off its rules directly.
Knowledge Cutoff
Date after which a model knows nothing from training, since its weights were frozen when training ended. RAG or a tool call like web search is how a system reaches past the cutoff.
Large Language Model (LLM)
AI trained on massive text datasets.
LLMs are a type of ML. ML is a type of AI.
Latency
How long the model takes to respond.
Loop
Repeated cycle in which a system observes a result, decides what to do next, acts, then feeds the outcome back in and starts again.
Agentic loop - Observe-think-act: the agent checks its own work, reacts, and refines until it reaches the goal or hits a stopping condition like a step limit or budget.
Reflection loop - Agent reviews its own output, spots problems, and revises; stops when the work is good enough or after a set number of passes.
Evaluation loop - Agent's output is scored against criteria, and the score drives the next attempt; stops when the output clears the bar or after a set number of tries.
Tool-use loop - Reason-act-observe: agent calls a tool, reads the result, then calls another or finishes when it has what it needs.
Human-in-the-loop - Person reviews, approves, or corrects the agent's work at set checkpoints; stops or continues based on their call.
Feedback loop - Any cycle where the result is measured and fed back to improve the next attempt; may run indefinitely unless given a target.
Training loop - Model predicts, gets feedback, and adjusts, repeated until it stops improving or a set number of passes is reached.
Machine Learning (ML)
Systems that learn patterns from data. Deep learning is a subset of ML using layered neural networks inspired by the brain.
LLMs are a type of ML. ML is a type of AI.
Memory
State an agent keeps beyond a single response, so it can carry information forward. Short-term memory is what's available within the current session, usually held in the context window. Long-term memory is persisted outside the context window in a file or database, and pulled back in when relevant, so an agent can recall things across sessions.
Model
Massive file of numbers (weights or parameters) that work together mathematically to generate text responses. The model has no memory, personality, or tools. It generates answers token by token based on probability.
Model Context Protocol (MCP)
Open standard that allows AI models to connect with external tools, data sources, and services to allow plug-and-play for tools. An MCP server exposes tools (model-controlled actions), resources (app-controlled context like files or DB rows), and prompts (user-invoked templates). By the time an MCP’s capabilities reach the model they have been flattened into ordinary tools to call.
Examples: search the web, access files, query databases, interact with apps like Slack or Gmail.
Multi-Agent
Systems where multiple AI models collaborate on a task.
Multimodal
Model that handles more than one type of input or output (text, images, audio, video) rather than just one. It folds jobs like computer vision and NLP into a single model, so it can read a screenshot and describe it, or take a voice note and answer in text.
Natural Language Processing (NLP)
AI understanding and generating human language.
Neural Network
Layered math structure, loosely inspired by the brain, that learns patterns by adjusting the strength of connections between its units. The architecture underneath deep learning, and the reason a model is ultimately just a large file of weights.
Orchestration
Coordinating multiple agents or steps so they work together toward one goal. The orchestrator is the layer that decides what runs when, in what order, and passes results between them. In a multi-agent setup, the orchestrator is usually itself an agent directing the others.
Persona
Defined character, identity, or personality assigned to a model to shape its tone, voice, and style of response. Overlaps with role prompting, but focuses on personality and communication style rather than expertise or task focus.
Prompt Engineering
Crafting inputs to get better AI outputs.
Standard prompting - Asking direct question or giving direct instruction in a short prompt.
Zero-shot prompting - Giving the model instructions with high specificity and low complexity without examples.
One-shot prompting - Including a standard pattern example in the prompt with the format or style of desired output. Shot means example, not attempt; it’s distinct from the slang "I one-shotted it" meaning solved in one try.
Few-shot prompting - Including a few diverse examples for complex cases, such as including a description, technical requirements, implementation and integration notes, and expected deliverable with data structure.
Role prompting - Assigning an identity or expertise for the tone or focus of the output.
Chain of thought prompting - Asking model to break down its reasoning into intermediate steps before giving a final answer (such as including “let’s think step by step” in a zero shot). Reasoning models now do this on their own, so the move is often setting how much they reason rather than including in prompt.
Emotional prompting - Including emotional language or stakes to influence the model's response.
Prompt chaining - Breaking a complex task into separate prompts where each output feeds into the next.
Negative prompting - Specifying what you don’t want in the output.
Meta-prompting - Having model write or improve prompts for itself or another model.
Prompt Injection
Attack where hidden instructions in content the model reads (web page, email, file, or tool result) try to hijack what it does. Injection exploits the fact that an agent acts on what it reads, slipping external content in as a command rather than data.
Reasoning Model
Model trained to work through a problem in steps before answering, spending more compute to think instead of replying immediately.
Retrieval Augmented Generation (RAG)
Giving an LLM the ability to look up relevant information from an external source before generating a response. Lets it stay up-to-date without retraining.
Skill
Modular set of instructions or best practices that improve AI output for a specific task. A skill extends what the model knows how to do; it's like a playbook or recipe that the model reads before starting a task. A skill can instruct the model to call tools, including tools that came from an MCP server.
Examples: write text documents, work with PDFs, create UI components, write tests, review pull requests.
Structured Output
Model returns its response in a specific, predictable shape that matches a provided template, schema, or data format, rather than free-form text.
Subagent
Agent spawned by another agent to handle a scoped piece of work and report its result back. Splitting a job across subagents keeps each one focused on a narrow task with its own context, instead of one agent juggling everything. The agent that spawns them coordinates the results.
System Prompt
Set of instructions given to the model before the user conversation begins that defines its role, behavior, and guidelines for how it should respond. In consumer AI products, the provider writes the system prompt; if you're building an application with an API, you write it yourself.
Temperature
Setting that controls how random a model's output is. Low temperature makes it pick the most likely next token and stay predictable. High temperature lets it wander into more varied and creative answers, at the cost of reliability.
Token
Units of text a model processes, used to determine cost to use. One token is about ¾ of a word. Input tokens (text given to model) are lower cost than output tokens (text generated by model).
Tool
A function that extends what the model can do; it's a new verb. Where a skill is consumed by being read, a tool is consumed by being called.
Training
Process for an AI model to learn from data to develop its capabilities. Billions of times, the model is educated by taking in data to make a prediction then using feedback to make better predictions. People fine-tune the model with curated data and use Reinforcement Learning from Human Feedback (RLHF) to produce preferred responses.
Vector Database
Storage built to hold embeddings and find the closest matches to a query quickly. When a RAG system looks something up, it's usually searching a vector database for the nearest embeddings, then feeding those results into the context window.
Workflow
System where the steps are fixed in advance by a person, and the model fills in each one, following a set path every time. A workflow is predictable, wired up ahead of the task like "summarize this, then translate it, then email it"; an agent is more flexible, figuring out its own steps from a goal like "get this translated and sent".
Worktree
A separate directory where a different branch of a code repository is checked out, letting you work on multiple branches at once without cloning the project each time. Individual agents can run in their own worktrees, making changes in isolation to avoid overwriting each other.
*AI was used for research