AgentCore course → Reference → Terminology
AgentCore terminology card
Built from Lesson 1. AgentCore's names are genuinely confusable. Several describe adjacent things, and two are plain English words used in a specific sense. This card exists so a name never derails you mid-meeting. All definitions verified 2026-08-05.
The distinctions people get wrong
Four pairs cause almost all the confusion. If you internalise only this section, the card has paid for itself.
| Often confused | The actual difference |
|---|---|
| Runtime vs harness | Runtime hosts an agent loop you wrote. The harness runs a loop AWS assembles from your configuration. Both are compute; you pick one. |
| Policy vs Guardrails | Policy decides whether an action is permitted (authorization). Guardrails inspects content for harm, injection, or PII. Policy can invoke Guardrails, but they ship, and are regionally available, separately. |
| Memory vs session isolation | Memory is what the agent remembers. Session isolation is a compute and security boundary in Runtime. A session can be perfectly isolated and remember nothing. |
| tools vs skills | A tool is a single callable capability reached through Gateway. A skill is a packaged bundle of capability an agent can access, including an AWS-curated catalog. Skills are a harness-era concept. |
Capabilities
- Runtime
- Serverless compute purpose-built to host an agent loop you wrote, with per-session isolation, fast cold starts, and support for long-running asynchronous agents. Avoid: "the AgentCore container", "the agent host"
- harness
- A managed agent loop defined in configuration rather than code (model, tools, skills, instructions) which AWS assembles and runs, each session in its own isolated environment with filesystem and shell. Avoid: "AgentCore Harness" (AWS lowercases it), "the managed agent"
- Gateway
- The capability that turns APIs, Lambda functions, and existing MCP servers into governed MCP tools reachable by an agent, and the single point every tool call passes through. Avoid: "tool proxy", "API gateway" (that is a different AWS service)
- Identity
- Agent identity and access management covering both inbound authentication (who is calling the agent) and outbound credentials (what the agent may call, and on whose behalf), against an existing identity provider. Avoid: "agent auth", "IAM for agents"
- Policy
- Deterministic authorization that intercepts every tool call at the Gateway perimeter before execution, authored in natural language or Cedar. Avoid: "guardrails", "agent permissions", "Policy" as a proper noun
- Memory
- Managed agent state, split into short-term memory for multi-turn conversation and long-term memory that persists across sessions and can be shared between agents. Avoid: "context store", "vector store", "checkpointer"
- Code Interpreter
- An isolated sandbox in which an agent can execute code it has written, in Python, JavaScript, or TypeScript. Avoid: "the sandbox" unqualified, since Browser is also a sandbox
- Browser
- A managed cloud browser runtime that lets an agent navigate sites, fill forms, and extract information, compatible with Playwright and Browser Use. Avoid: "headless browser", "web tool"
- Built-in Tools
- AWS's collective name for Browser and Code Interpreter together. It appears in the region availability table but rarely elsewhere. Avoid: using it where you mean one specific tool
- Observability
- OpenTelemetry-native tracing of every step in an agent's execution path, including intermediate outputs, surfaced through Amazon CloudWatch. Avoid: "logging", "monitoring". The step-level trace is the point
- Evaluations
- Automated assessment of agent quality over sessions, traces, and spans, using built-in or custom evaluators. Avoid: "evals" in customer-facing writing; "testing"
- Optimization
- Trace-driven recommendations for improving system prompts and tool descriptions, validated by batch evaluation or A/B traffic splitting. Avoid: "tuning", "fine-tuning". No model weights change
- AWS Agent Registry
- An organisation-wide catalog of agents, MCP servers, tools, and skills with an approval workflow and combined semantic and keyword search. Avoid: "AgentCore Registry", which is namespaced separately from AgentCore
- Payments
- Managed agent micropayments over the x402 protocol, with third-party wallets and deterministic per-session spending limits. Avoid: "agent billing", which means your bill, not the agent's spending
Protocol and pricing vocabulary
- MCP (Model Context Protocol)
- The open protocol AgentCore uses to expose tools to agents. Gateway's output format and Registry's record types both speak it.
- A2A (Agent-to-Agent)
- The open protocol for agents to discover and delegate to other agents.
Supported by Runtime and catalogued by Registry as an
a2aAgentCard.
- x402
- The payment protocol built on HTTP 402. An agent hits a paid resource, receives 402, and Payments negotiates and settles without breaking the reasoning loop.
- Cedar
- AWS's open-source Policy language, one of the two authoring options for Policy alongside natural language.
- Cross-region inference (CRIS)
- Automatic routing of a model inference request to another region for capacity. Applies to Memory, Policy, and Evaluations. Data is stored only in the primary region, but prompts and outputs may leave it. See the Canada card.
- Memory event
- One unit of short-term memory written, and one of Memory's three billing meters. Distinct from a memory record, which is a stored long-term item.
- Memory strategy
- How long-term memory is derived and stored: built-in, built-in with overrides, or self-managed. The choice changes both cost and whether inference crosses regions.
- Session
- In this workspace, always a Runtime or harness execution session, meaning the isolation boundary. Not a user's login session, and not the lifetime of a conversation.
Naming trap worth remembering
AWS writes harness, Policy, and payments in lower case, but Runtime, Memory, Gateway, Identity, and Evaluations capitalised. It reads like inconsistency because it is. The lower-cased ones are described as capabilities within AgentCore rather than standalone services. No customer will care, but getting it backwards makes you look careless in written follow-ups.