Visual overview
Bedrock/AgentCore supplies model-driven orchestration, IAM constrains AWS authority, and Lambda can implement controlled business actions.
Technical reference
Agent services evolve quickly. These mechanics are the durable control plane; verify the current AgentCore capability status and model/tool support before implementation.
The model proposes steps while the runtime executes tools and returns observations that inform later decisions.
Precise descriptions and parameter contracts help tool selection, but application-side validation is still mandatory.
AgentCore Gateway can expose supported services and APIs through MCP-compatible tool interfaces.
Retention scope should match privacy and product requirements rather than storing every interaction indefinitely.
Tool calls should inherit only permissions justified by the user, agent, and requested action.
Observability is essential for debugging loops, measuring task success, detecting failures, and auditing consequential workflows.
Service limits and capabilities can change. Values shown here reflect the current AWS documentation; use the linked official sources below as the source of truth.
An agent couples model reasoning with an execution loop
A generative model call usually maps input context to output. An agent adds a loop: interpret a goal, choose a step, optionally invoke a tool, observe the result, update context, and decide whether another step is needed. Tools can expose functions, APIs, databases, search, code execution, or business systems. Retrieval supplies information, while tool use can create side effects such as opening a ticket, updating a record, or triggering an operational workflow.
Tool calling remains an application contract. The model can propose a tool and arguments, but the surrounding runtime should validate the tool name, schema, parameters, authorization, and result before execution continues. A strong system separates reasoning from authority: the model does not become more trusted merely because it selected an action confidently. High-impact or irreversible operations can require deterministic policy checks or explicit human approval.
AgentCore provides modular infrastructure for modern agent systems
Amazon Bedrock AgentCore is a managed platform for deploying and operating agents built with different frameworks and model providers. Its modular services include runtime/harness capabilities, memory, Gateway for exposing tools and APIs including MCP-compatible interfaces, identity, observability, code and browser tools, evaluation, and policy capabilities. These pieces can be adopted independently, so an architecture does not have to move every concern into one monolithic agent service.
Identity and policy deserve special attention. An agent can act for different users and can access third-party systems, so static shared credentials undermine least privilege and auditing. AgentCore Identity is designed for agent and workload authentication/credential flows, while Gateway and policy controls can constrain tool access. Observability exposes sessions, traces, spans, and service metrics so teams can inspect which steps and tools contributed to an outcome.
Know the lifecycle status of Bedrock Agents Classic
Amazon Bedrock Agents Classic remains important for understanding the orchestration pattern. A classic agent combines instructions, action groups, optional knowledge bases, and a foundation model. During orchestration it can predict an action, invoke a Lambda-backed function or return control to the application, observe the result, query a knowledge base, and repeat. Trace output exposes the sequence for debugging and evaluation.
However, AWS documentation states that Bedrock Agents Classic stopped accepting new customers on July 30, 2026; existing customers can continue to use it. For new architectures, AWS directs builders toward Amazon Bedrock AgentCore for similar agent infrastructure. This distinction is time-sensitive: learn the durable concepts—tools, observations, memory, identity, orchestration and policy—while checking current AWS availability before choosing the concrete agent service.
Key takeaways
- 01
Agents add iterative planning/action/observation around model inference.
- 02
Tool selection by a model does not grant authority; applications must validate schemas, permissions, and side effects.
- 03
AgentCore provides modular runtime, memory, Gateway, identity, observability, evaluation, and policy capabilities for agent systems.
- 04
MCP can standardize how agents discover and call external tools, but authorization remains an application/platform responsibility.
- 05
Bedrock Agents Classic is no longer open to new customers as of July 30, 2026; existing customers can continue using it.
Official AWS sources
Use these primary AWS resources for the source material behind this article and for deeper reference.