Glossary

A–Z reference for every term used in the course. Each entry links back to the module that introduces it.

Agent
A program where an LLM decides the next step (often a tool call) in a loop until the task is done. Module 1.
AgentExecutor
Legacy LangChain helper to run a ReAct loop. Superseded by LangGraph. Module 2.
add_messages
The LangGraph reducer that appends new messages to the existing list (rather than replacing). Module 4.
AIMessage
A message authored by the model. May contain text and/or tool calls. Module 3.
Assistant
In LangGraph Platform: a graph paired with a specific configuration (prompt, model, flags). Multiple assistants can share a graph. Module 9.
Breakpoint
A node where execution pauses; declared at compile time via interrupt_before / interrupt_after. Module 8.
Checkpointer
The component that persists graph state after every step, keyed by thread_id. SQLite for dev, Postgres for prod. Module 6.
Command(resume=…)
Used to continue a paused run after an interrupt(), passing the human-provided value back into the graph. Module 8.
Conditional edge
An edge whose destination is decided at runtime by a function. Module 4.
Configurable
Values passed via config.configurable that the LLM never sees — user identity, tenant id, feature flags. Module 5.
Control plane
The management surface (UI + API) where you deploy, configure, and observe — separate from the data plane that runs traffic. Module 9.
Data plane
The component that actually executes user runs — API server, workers, queue, Postgres. Module 9.
Edge
A wire from one LangGraph node to another. Regular or conditional. Module 4.
Episodic memory
Long-term memory of past events / experiences for a user. Module 6.
Evaluator
A function that grades a run's output. Heuristic / LLM-as-judge / human. Module 10.
FakeListChatModel
A test-only model that returns a pre-canned list of responses; used to unit-test graphs without paying for tokens. Module 11.
Guardrail
An automated input / output / tool check that runs when no human is in the loop. Module 8.
Handoff
A tool one agent calls to transfer active control to another agent (swarm pattern). Module 7.
Hierarchical pattern
Supervisors all the way down: a top supervisor routes to mid-level supervisors, which route to specialists. Module 7.
HumanMessage
A message authored by the user. Module 3.
Interrupt
The LangGraph primitive that pauses a run mid-execution, surfaces a payload, and waits for resume. Module 8.
LangChain
The building-blocks library: chat models, messages, prompts, tools, retrievers, parsers. Module 2.
LangGraph
The orchestration library: state + nodes + edges for building stateful, multi-step, possibly cyclic LLM apps. Module 2.
LangGraph Platform
The deploy/manage layer: control plane + managed data plane (assistants, threads, runs, scaling). Module 9.
LangSmith
The observability + evaluation product: traces, datasets, evaluators, online monitoring. Module 10.
LLM-as-judge
An evaluator that uses another LLM to grade outputs against a rubric. Module 10.
MCP (Model Context Protocol)
Open standard for tool servers any MCP-aware client can connect to. Module 5.
Multi-agent system
A system of multiple cooperating agents — supervisor, network, hierarchical, or swarm. Module 7.
Multitask strategy
How to handle a new user message while a run is in progress: reject / enqueue / interrupt / rollback. Module 9.
Network pattern
Multi-agent shape where any agent can hand off to any other; no central boss. Module 7.
Node
A function in a LangGraph: takes state, returns a partial state update. Module 4.
Online eval
An evaluator that runs continuously against a sample of production traffic. Module 10.
Offline eval
An evaluator run pre-deploy against a fixed gold dataset, often in CI. Module 10.
Procedural memory
Long-term memory of how to do things — playbooks, runbooks. Module 6.
Prompt injection
Adversarial instructions hidden in untrusted text (emails, retrieved docs) that try to hijack the agent. Module 11.
RAG (Retrieval-Augmented Generation)
Fetching relevant chunks from a vector store and feeding them to the LLM. In agents, exposed as a tool. Module 6.
ReAct
The "reason → act → observe" loop most agents follow. Module 1.
Reducer
A function that combines an old state-field value with a new one returned by a node. Module 4.
Run
One invocation of an assistant on a thread. Lives for seconds to minutes. Module 9.
RunnableConfig
The config dict carried through a graph; holds configurable, thread_id, callbacks, tags. Module 5.
Semantic memory
Long-term memory of facts about the world / user. Module 6.
StateGraph
LangGraph's main class for defining graphs around a typed state. Module 4.
Store
Cross-thread key-value persistence layer for long-term memory, with optional vector index. Module 6.
Structured output
Asking the model to return a typed object matching a Pydantic / Zod schema. Module 3.
Subgraph
A compiled graph used as a node inside a larger graph. Module 7.
Supervisor pattern
Multi-agent shape where one orchestrator routes each turn to a specialist. Module 7.
Swarm pattern
Multi-agent shape where active agents hand off to each other directly. Module 7.
SystemMessage
The instructions / persona prompt at the start of a conversation. Module 3.
Thread
A persistent conversation; one checkpointed state keyed by thread_id. Module 6.
Time travel
Rewinding a thread to a past checkpoint and (optionally) editing state before resuming. Module 8.
Tool
A typed function the LLM can request to call. Defined by name + description + schema + body. Module 5.
ToolMessage
A message holding the result of a tool execution. Tied to the requesting tool call by id. Module 3.
ToolNode
The LangGraph prebuilt node that executes whatever tool calls the LLM emitted, in parallel. Module 5.
Trace
A tree of every LLM/tool/node call in one run, captured by LangSmith. Module 10.
Vector store
A database that indexes embeddings for similarity search (pgvector, Chroma, Qdrant, …). Module 6.
Workflow (vs. agent)
A pre-scripted sequence of steps. Choose over an agent when the path is knowable in advance. Module 1.