Glossary

Foundational LLM and Agent Concepts

Term	Definition
Large language model (LLM)	A statistical model trained on huge amounts of text that generates responses by predicting the most likely next chunk of text given what came before.
—	—
Token	The unit an LLM works in — roughly three-quarters of a word, or a few characters of code — used to measure context, cost, and speed.
—	—
Context window	The total amount of text (measured in tokens) that a model can consider at once, including your prompt, prior turns, and any files it has read.
—	—
Pre-training	The initial training phase that makes a raw LLM fluent at predicting text, but does not yet make it useful as an assistant.
—	—
Post-training	The subsequent training phases (SFT, RLHF, tool-use fine-tuning) that turn a fluent LLM into a helpful, instruction-following assistant.
—	—
Supervised fine-tuning (SFT)	A post-training phase in which the model learns from human-written examples of ideal user-and-assistant exchanges.
—	—
Reinforcement learning from human feedback (RLHF)	A post-training phase in which the model learns from human preference judgments about which of two responses is better.
—	—
Tool-use fine-tuning	A post-training phase that teaches the model when and how to call external tools rather than just talking about them.
—	—
Transformer	The neural network architecture used by all modern LLMs, introduced by Google researchers in 2017.
—	—
Attention	The mechanism inside a Transformer that lets the model weigh the relevance of every part of its input when generating each next token — the reason context engineering works at all.
—	—
Chat interface	A conversational interface to an LLM where the model responds but cannot take actions in the world.
—	—
Agent	An LLM plus a surrounding system that lets it take actions — reading files, running commands, calling APIs — in a loop, until a task is done.
—	—
Agent loop	The cycle of sending a conversation to the model, executing any tools it requests, feeding results back, and repeating until the model has no more tool calls to make.
—	—
Harness	The software wrapper around an LLM that manages tools, memory, permissions, and the agent loop; the same model behaves differently in different harnesses.
—	—
Interface	What the researcher actually sees and interacts with — a chat window, a command-line tool, an IDE panel — as distinct from the model and the harness underneath.
—	—
System prompt	An invisible instruction the harness prepends to every conversation to shape the model’s behavior.
—	—
Statelessness	The fact that models have no memory between conversations; anything that looks like memory is the harness re-loading context on each turn.
—	—
Convergent post-training	The observation that different labs (Anthropic, OpenAI, Google, Meta) use similar training pipelines and their models therefore behave similarly, which is why swapping models mostly works.
—	—

Context, Skills, Rules, and Related Authoring Surfaces

Term	Definition
Context engineering	The practice of deliberately shaping what information the agent has access to within its context window, in order to control the quality of its output.
—	—
Prompt	The text you send to the model, including instructions, examples, and any context you want it to consider.
—	—
Project memory	A file the harness automatically loads at the start of every session so the agent knows about your project without being retold.
—	—
AGENTS.md	The emerging cross-vendor convention for the project memory file, containing project structure, conventions, and domain context, like a README for the project
—	—
CLAUDE.md	Claude Code’s specific project memory file; functionally equivalent to AGENTS.md in other tools. For other platforms: GEMINI.md
—	—
Skill	A reusable, named procedure the agent can invoke on demand when a task matches its description.
—	—
SKILL.md	The markdown file convention for defining a skill, with a description at the top that acts as the agent’s selector and a body containing the procedure.
—	—
Rule	A path-scoped constraint that applies when the agent is editing files matching a specific pattern.
—	—
Subagent	A specialized version of the agent with a distinct persona and often restricted tool access, used when different permissions or a different voice are needed.
—	—
Slash command	A user-typed shortcut (like /clear or /init) that triggers a specific harness behavior.
—	—
Progressive disclosure	The design principle of showing a short summary up front and revealing full detail only when it becomes relevant — how skills and nested context files are structured.
—	—
Handoff artifact	A compacted document that summarizes a session’s state so a fresh session can pick up where the last one left off.
—	—

Model Context Protocol (MCP)

Term	Definition
Model Context Protocol (MCP)	A standardized protocol that lets agents connect to external tools and data sources without custom integration per vendor.
—	—
MCP server	A separate service that exposes tools, resources, and prompts through the MCP protocol so any MCP-compatible agent can use them.
—	—
MCP client	The side of the protocol that lives inside the agent’s harness and talks to MCP servers on the agent’s behalf.
—	—
MCP tool	A callable function exposed by an MCP server that the agent can invoke — one of the three MCP primitives.
—	—
MCP resource	Structured data exposed by an MCP server that the agent can read — a second MCP primitive.
—	—
MCP prompt	A reusable prompt template exposed by an MCP server, often used to guide the agent’s workflow — the third MCP primitive.
—	—
Tool description	The natural-language explanation of what a tool does, which the agent reads to decide whether to call it — arguably the most important piece of MCP server design.
—	—
Streamable HTTP transport	The MCP transport mechanism used by remotely hosted servers, allowing agents to communicate with them over HTTP.
—	—
STDIO transport	An MCP transport mechanism where the server runs as a local process and communicates through standard input/output — common for local development.
—	—
MCP Inspector	A browser-based tool for testing MCP servers by connecting to them, listing their tools, and invoking them interactively.
—	—

Sandboxing, Security, and Permissions

Term	Definition
Sandbox	An environment that hosts a workspace and limits what an agent can touch — restricting file access, network calls, and command execution to a defined boundary.
—	—
Proceed-in-sandbox	The pattern of auto-approving agent actions inside a sandbox boundary while requiring confirmation to cross it.
—	—
Approval habit	The discipline of reading what the agent proposes to do before accepting it, especially for file edits and shell commands.
—	—
Prompt injection	An attack in which instructions hidden inside content the agent reads (a file, a webpage, a data cell) hijack the agent’s behavior.
—	—
Read-only skill	A skill whose prose constraints tell the agent to inspect but not modify — a soft, description-based constraint on behavior.
—	—
Credential hygiene	The practice of keeping secrets, tokens, and passwords out of any context an agent might see.
—	—

Data, Privacy, and Institutional Practice

Term	Definition
Consumer tier	The free or personal-paid access to a model where prompts are often logged and may be used to train future models.
—	—
Enterprise tier	Paid institutional access to a model with contractual guarantees that prompts are not logged or used for training.
—	—
API access	Programmatic access to a model, typically billed by token, often with the same privacy guarantees as enterprise tiers.
—	—
Local model	A model running entirely on your own hardware, so no data leaves your machine at the cost of typically lower capability.
—	—
LLM gateway	An institution-level proxy that routes agent requests through a controlled endpoint for authentication, budget enforcement, and audit logging.
—	—
Redaction	The practice of stripping sensitive fields from data before sending it to an agent so the agent can help without seeing the sensitive parts.
—	—
Synthetic sample	A fabricated example that preserves the shape of real data, used in place of real data when the actual values are sensitive.
—	—

Verification, Testing, and Reproducibility

Term	Definition
Verification	The process of checking that an agent’s output is actually correct, especially important because plausible-sounding outputs can be silently wrong.
—	—
Eval	A scored test suite that measures how well an agent or skill performs on a domain-specific task, distinct from unit tests of the code the agent writes.
—	—
LLM-as-judge	The pattern of using one model to evaluate another model’s output against a rubric — the mechanism behind most modern eval pipelines.
—	—
Hallucination	A confident-sounding but incorrect or fabricated output — a common failure mode, especially in domains underrepresented in training data.
—	—
Provenance chain	The trail from prompt to AI response to human review to committed artifact that documents how AI-assisted work was produced.
—	—
Auditability trail	Per-phase markdown artifacts (plans, analyses, validations) committed alongside code that document why decisions were made.
—	—
Determinism	The property of getting the same output for the same input; not a default for LLMs but achievable via API parameters like temperature and seed.
—	—
Model drift	The phenomenon of a model producing different outputs over time as vendors update the underlying model, breaking reproducibility across time.
—	—

Retrieval and Grounding

Term	Definition
Retrieval-augmented generation (RAG)	An approach that gives the model access to your specific documents by retrieving relevant chunks and inserting them into the prompt.
—	—
Embedding	A numerical representation of text that captures its meaning in a way that can be compared to other text for similarity.
—	—
Vector database	A specialized database that stores embeddings and lets you efficiently find text similar to a query.
—	—
Chunking	The process of splitting long documents into smaller pieces that can be individually retrieved and inserted into a prompt.
—	—
Semantic search	A search approach that finds documents by meaning rather than exact word matching, powered by embeddings.
—	—
Knowledge cutoff	The date after which the model was not trained on new information, meaning it does not know about later events.
—	—
Grounding	Anchoring an agent’s response in specific documents or data sources you provide, rather than relying on its training knowledge.
—	—

Workflow Patterns

Term	Definition
Research/Plan/Implement (RPI)	A workflow pattern that separates a task into a research phase, a planning phase, and an implementation phase, with an auditable artifact after each.
—	—
Explore/Plan/Execute (EPE)	A variant of RPI with slightly different phase names, used in some ecosystems.
—	—
Spec-driven development	Writing a specification document before writing code and using the agent to critique and refine it.
—	—
Vibe coding	The informal pattern of prompting an agent conversationally and accepting what it produces without structured discipline — contrasted with workflow-driven approaches.
—	—
Trained in vs. in the prompt	The mental model that distinguishes what the model already knows from training (which you can’t change) from what you provide in the current session (which is your only lever).
—	—

Bias, Ethics, and Failure Modes

Term	Definition
Sycophancy	The tendency of models to agree with the user’s framing rather than push back, a systematic side-effect of preference-based training.
—	—
Context exhaustion	The failure mode where an agent forgets earlier parts of a conversation as the context window fills up.
—	—
Confident wrong answer	The failure mode where an agent produces a plausible-sounding but incorrect result, made worse by RLHF training that rewards confident-sounding responses.
—	—
Reward hacking	A trained-in tendency for models to produce outputs that score well on their training reward without actually being helpful or correct.
—	—
Scope creep	The failure mode where an agent does more than it was asked, a consequence of training that rewards being helpful.
—	—