Every other chapter uses these words. Read this once and the rest will make sense. Each term has a one-line meaning and, where it helps, a simple comparison.
Agent
A program that uses an AI model to reach a goal. It can decide what to do, call tools, and use the results, not just reply with text.
Modelthe LLM
The AI "brain" (for example Anthropic Claude) that reads, reasons, and writes. It does not act by itself. The agent gives it tools.
Tool
Anything the agent can call to do real work or fetch real data: a function you wrote, an API, a database query, a web search. Think of tools as the agent's hands.
Agent framework
A code library for building agents: Strands, LangGraph, CrewAI, LlamaIndex, Google ADK. It wires the model, the instructions, and the tools together.
The agentic loop
The repeating cycle an agent runs: the model thinks, calls a tool, reads the result, thinks again, and so on, until it has a final answer.
Session
One continuous run or conversation, named by a session id. Like a single phone call. Reusing the id continues the same conversation.
Serverless
You run and manage no servers. AWS starts, scales, and stops the compute for you, and you pay only for what you use.
Container image
A sealed package of your code plus everything it needs to run. AgentCore can run your agent from a container image, or from directly deployed code (a zip).
ARM64AWS Graviton
The processor type AgentCore runs on. Your container image must be built for ARM64, or it will not start.
microVM
A tiny, private virtual computer created for one session, with its own CPU, memory, and disk. One user's data cannot leak into another's. It is destroyed and its memory wiped when the session ends (by default; optional session storage can keep files). Like a private, single-use room.
/ping health check
A small status endpoint AgentCore calls to ask "are you healthy, and are you busy?" If an agent doing long work does not answer "busy," it is treated as idle and stopped.
Session replay
A recording of a Browser session (pages visited, clicks, actions) you can watch back later, like a screen recording, to see exactly what the agent did.
MCPModel Context Protocol
An open standard for how an agent finds and calls tools. Like a universal plug (think USB) so any agent can use any MCP tool.
A2AAgent-to-Agent
An open standard for how one agent talks to another agent to hand off or share work, even across different companies. A shared language for agents to collaborate.
Namespace
A folder-like path that organizes an agent's long-term memories, so they are easy to find, filter, and control.
Gatewayan AgentCore service
Turns your existing APIs and AWS Lambda functions into MCP tools, behind one secure address, so agents can use them without custom glue code.
IAM
AWS's permission system: who or what may do which actions. Least privilege means granting only the access actually needed.
OAuth scope
The exact permissions a login token grants (for example, "read your calendar"). Too broad a scope hands over more access than necessary.
KMSKey Management Service
AWS's service for encryption keys. A customer-managed key means you own and control the key that encrypts your data.
Trace and spanobservability
A record of what the agent did. A trace is one whole request. A span is one step inside it (a model call, a tool call). Like a flight recorder for debugging.
x402
An open web payment standard (it reuses the HTTP "402 Payment Required" response) that lets software make tiny automatic payments for paid services. Note: this is real money, not the "tokens" an LLM processes.
Multi-modal
Able to handle more than text: images, audio, files, not just words.