Amazon Bedrock Agents

Bedrock Agents decomposes user requests into multi-step workflows by analyzing intent and automatically selecting which actions (Lambda functions) to invoke in sequence. The agent maintains state across steps, evaluates intermediate results, and determines next actions without explicit step-by-step programming. Uses foundation model reasoning to map user goals to action chains, with built-in loop detection and termination logic.

Bedrock Agents invoke Lambda functions as 'action groups' by matching agent-selected actions to Lambda endpoints via OpenAPI-style schemas. Each action group defines input/output schemas that the agent uses to construct Lambda payloads and interpret responses. The agent automatically maps its reasoning to the correct Lambda function and parameter binding without manual routing logic.

Bedrock Agents support streaming responses where results are returned incrementally as the agent reasons and executes actions, rather than waiting for complete execution. Streaming enables real-time feedback to users and reduces perceived latency. Supports both event-stream and chunked transfer encoding for streaming responses.

Bedrock Agents integrate with AWS CloudWatch and X-Ray for monitoring agent invocations, tracking latency, action execution, and error rates. Provides metrics on agent reasoning steps, action invocations, and guardrail violations. Enables debugging of agent behavior through execution traces and logs without custom instrumentation.

Bedrock Agents connect to knowledge bases (document collections indexed in Amazon Kendra or OpenSearch) to retrieve relevant context before generating responses. The agent automatically decides when to query the knowledge base, constructs retrieval queries from user intent, and augments its reasoning with retrieved documents. Supports semantic search and keyword matching across structured and unstructured data.

Bedrock Agents maintain conversation history within a session, allowing multi-turn interactions where the agent retains context from prior exchanges. Session state is managed server-side by Bedrock, with automatic context windowing to fit within foundation model limits. Agents can reference prior user intents and action results without explicit memory management by the caller.

Bedrock Agents apply guardrails (configurable safety policies) to filter harmful content, enforce topic boundaries, and prevent misuse. Guardrails intercept both user inputs and agent outputs, checking against predefined or custom filters for toxicity, PII, off-topic requests, and policy violations. Violations trigger configurable responses (block, redact, or alert) without invoking agent reasoning.

Bedrock Agents can pause execution and return control to the caller at specified decision points, allowing human review or approval before proceeding. The agent provides context (reasoning, proposed actions, intermediate results) and waits for human input to continue. Enables workflows where high-stakes decisions require human judgment before agent action execution.

Bedrock Agents run on foundation models available in Amazon Bedrock (Claude, Llama, Mistral, etc.), with caller ability to select which model powers the agent. Different models have different reasoning capabilities, latency profiles, and cost characteristics. Agents automatically adapt to model capabilities (e.g., token limits, tool-calling support) without code changes.

Bedrock Agents can interpret and execute code (Python, JavaScript, etc.) as part of multi-step workflows, enabling agents to perform calculations, data transformations, or logic execution without invoking external services. Code execution is sandboxed and scoped to agent session, with results fed back into agent reasoning for subsequent steps.

Bedrock Agents are configured with custom instructions (system prompts) that define agent behavior, task scope, constraints, and reasoning style. Instructions control what the agent can do, how it makes decisions, and how it interacts with users. Supports dynamic instruction injection per invocation for context-specific behavior without redeploying the agent.

Bedrock Agents support multi-agent architectures where a supervisor agent coordinates multiple specialized agents, each handling specific domains or tasks. The supervisor agent decides which specialized agent to invoke based on user intent, aggregates results, and synthesizes final responses. Enables decomposition of complex problems into specialized sub-agents.