Glossary
This glossary defines the terms used across the guide. Many disagreements about AI engineering tools are really vocabulary problems.
AI-DLC
AI-Driven Development Life Cycle: a software delivery lifecycle where AI participates in requirements, design, planning, implementation, testing, documentation, and operations while humans retain accountability.
Do not confuse with: generic code generation or a single coding assistant.
Spec-Driven Development
A development approach where the specification is the source of truth and implementation follows from explicit behavior, constraints, acceptance criteria, and non-goals.
Do not confuse with: waterfall. Specs can be iterative and lightweight.
Workflow framework
A framework that defines how humans and agents move from intent to delivery artifacts such as specs, plans, tasks, approvals, tests, reviews, and audit records.
Do not confuse with: an app framework that runs production AI behavior.
Agent harness
The environment that runs an agent with access to tools such as filesystem, shell, Git, editors, memory, and subagents.
Do not confuse with: the workflow methodology that tells the agent what process to follow.
Agent runtime
The execution system for an agent, including state, tools, memory, model calls, and control flow. In coding contexts, this often overlaps with the harness.
Do not confuse with: the model itself.
Agent app framework
A framework for building AI applications or agent services, such as chains, graphs, tools, memory, state, and deployment surfaces.
Do not confuse with: a coding workflow like Spec Kit or AI-DLC.
Orchestration
The coordination of model calls, tools, state transitions, retrieval, memory, and human handoffs.
Do not confuse with: governance. Orchestration runs behavior; governance controls risk and accountability.
LangChain
An application framework for building LLM-powered apps and agents with integrations for models, tools, retrievers, prompts, and observability.
Do not confuse with: a lifecycle framework for software delivery.
LangGraph
A framework/runtime for building stateful, long-running, graph-based agent systems with durable execution and human-in-the-loop patterns.
Do not confuse with: a spec or governance framework.
Hermes
An open agent harness/runtime that can be used to run agents with tools, memory, skills, and subagent-style capabilities.
Do not confuse with: managed coding CLIs or enterprise lifecycle governance.
MCP
Model Context Protocol: a protocol for exposing tools, resources, and prompts to AI applications and agents in a standardized way.
Do not confuse with: a full security/governance system. MCP needs policy and controls around it.
Tool calling
The mechanism where a model produces structured arguments for a tool or function that an application may execute.
Do not confuse with: permission to execute every action. Tool calling still needs authorization.
Tool gateway
An organizational control layer that mediates tool access, authentication, authorization, rate limits, approval, logging, and audit.
Do not confuse with: MCP itself. A gateway can expose MCP tools, OpenAPI tools, scripts, and internal services.
RAG
Retrieval-Augmented Generation: a pattern where relevant external information is retrieved and added to model context before generating an answer or action.
Do not confuse with: simply storing embeddings in a vector database.
Embedding
A numeric representation of text, image, or other content used for semantic similarity search and retrieval.
Do not confuse with: the original content. Embeddings are an index representation.
Vector database
A database or index optimized for storing vectors and searching by similarity.
Do not confuse with: a complete RAG system. It is only one component.
Retriever
The component that selects candidate context from an index or source system for a query.
Do not confuse with: the generator model that writes the final response.
Reranker
A component that reorders retrieved candidates to improve relevance before they enter the prompt context.
Do not confuse with: the initial retriever. Reranking usually happens after retrieval.
Grounding
The practice of making model responses rely on explicit evidence, context, citations, or tool results.
Do not confuse with: model confidence. A confident answer can still be ungrounded.
Evaluation
The measurement of AI behavior against expected outcomes, including retrieval quality, generation quality, tool trajectories, safety, latency, and cost.
Do not confuse with: unit testing only.
Golden dataset
A curated set of representative cases with expected behavior or evidence used to detect regressions.
Do not confuse with: random demo prompts.
Trace
A recorded timeline of an AI run, including prompts, model calls, retrieval, tools, outputs, errors, and metadata.
Do not confuse with: plain application logs. Traces show end-to-end causality.
Observability
The ability to understand AI system behavior through traces, metrics, logs, evals, costs, latency, and feedback loops.
Do not confuse with: dashboards alone.
Guardrail
A control that prevents, detects, or mitigates unsafe or undesired AI behavior.
Do not confuse with: a guarantee. Guardrails reduce risk but must be tested.
Model router
A policy layer that routes requests to different models based on task, cost, latency, sensitivity, capability, or availability.
Do not confuse with: a single model endpoint.
Local LLM
A language model served inside a local, private, or self-controlled environment.
Do not confuse with: automatically safer or cheaper. It still needs evals, security, and operations.
Human-in-the-loop
A design where a human reviews, approves, corrects, or guides an AI action before or during execution.
Do not confuse with: rubber-stamp approval.
Audit trail
A durable record of decisions, approvals, tool calls, model runs, and evidence used to explain what happened.
Do not confuse with: chat history that can be lost or edited.
Approval gate
A point in the workflow where progress requires explicit approval from a responsible person or system.
Do not confuse with: informal acknowledgement in chat.
Source of truth
The authoritative artifact that defines current expected behavior, such as a spec, change proposal, audit record, or production configuration.
Do not confuse with: scattered notes across tickets, chats, and PR descriptions.
Context rot
The degradation of agent performance when long conversations accumulate outdated, irrelevant, or conflicting context.
Do not confuse with: model weakness alone. Better artifacts and session boundaries can reduce context rot.