Glossary — AWS Strands

A

Agent Loop

The core cycle in Strands where messages are sent to the model, tool_use responses are executed, results are appended, and the loop repeats until the model signals end_turn or a limit is reached.

Agent SOP

A markdown document using RFC 2119 keywords (MUST, SHOULD, MAY) that defines a structured workflow for an agent, providing deterministic process control while preserving the model's reasoning ability within each step.

AppState

A persistent key-value store on the agent (agent.appState) that is not sent to the model, used for cross-invocation data like counters, preferences, and accumulated results.

B

Bedrock AgentCore

Amazon's managed hosting service for Strands agents that handles scaling, monitoring, session management, and infrastructure automatically.

BedrockModel

The model provider class in Strands that handles streaming, token counting, and retries for communication with AWS Bedrock foundation models.

C

Context Offloader

A plugin that automatically moves large tool results to external storage (memory or S3) and provides a retrieval tool so the agent can fetch them back on demand, preventing context window exhaustion.

ConversationManager

An abstraction that controls how message history grows, with built-in implementations including NullConversationManager (no management), SlidingWindowConversationManager (keeps N most recent), and SummarizingConversationManager (compresses old messages into summaries).

G

Graph

A multi-agent orchestration pattern where developer-defined nodes and directed edges form a topology, the LLM at each node decides which edge to follow, and cycles are allowed for review loops.

H

Hooks

Lifecycle interceptors that fire at seven well-defined points in the agent loop (BeforeInvocation, BeforeModelCall, AfterModelCall, MessageAdded, BeforeToolCall, AfterToolCall, AfterInvocation), enabling cross-cutting concerns like logging, rate-limiting, and guardrails without modifying agent logic.

I

InvocationState

A per-invoke ephemeral context object shared by reference across all hooks and tools within a single invoke() call, used for request-scoped data like requestIds or accumulated tool results.

M

MCP (Model Context Protocol)

A protocol that allows tools written in any language to be exposed to Strands agents via stdio or HTTP transports, enabling polyglot tool implementations.

P

Plugin

A composable extension that bundles related hooks and optionally provides tools, implementing the Plugin interface with initAgent() for setup and lifecycle methods for observation and control.

S

SessionManager

A component that snapshots full agent state (messages, appState, conversation manager state) to pluggable storage backends (FileStorage, S3Storage, or custom), enabling multi-turn conversations across process restarts.

Steering Plugin

A built-in plugin that injects context-aware guidance fragments into the system prompt based on conditional rules, enabling modular, testable, and reusable prompt composition.

Stop Reason

The enumerated reason why an agent loop terminated, including endTurn (normal), limitTurns, limitTotalTokens, limitOutputTokens, cancelled, contentFiltered, and guardrailIntervention.

Structured Output

A feature where a Zod schema on the agent constructor constrains the model's final response to a validated, typed shape, with automatic retries when Zod refinements fail.

Swarm

A multi-agent orchestration pattern where agents in a pool have auto-injected handoff tools to autonomously transfer control to peers, creating emergent execution paths without developer-defined edges.

T

Tool (tool() function)

The factory function for creating custom tools in Strands TypeScript, accepting a name, description, Zod inputSchema (which generates JSON Schema for the model and validates input at runtime), and an async callback.

ToolContext

The optional second parameter in a tool callback providing access to the running agent instance (appState, cancelSignal, messages), current tool use metadata (toolUseId, name), and per-invocation ephemeral state.

W

Workflow

A multi-agent orchestration pattern defining a fixed DAG of tasks with explicit dependencies, where independent tasks run in parallel and execution order is fully deterministic.