Building and Deploying Agents
Build and deploy agents by defining goals, instructions, reasoning models, and tools. Configure Amazon Bedrock Agents or AgentCore, authorize tool access, enable agent communication through MCP and A2A protocols, and coordinate specialized agents with a supervisor.
We'll cover the following...
A single prompt returns a single answer. An agent receives a goal, such as “refund order 4521 if it is eligible,” and works toward it in a loop: the foundation model reasons about the next step, calls a tool, reads the result, and decides whether to call another tool or answer. The Bedrock Agents and Guardrails lesson in Chapter 8 introduced this ReAct-style loop with Amazon Bedrock Agents. This lesson turns that concept into deployment decisions: which tools the agent gets, how it reaches them, how it talks to other agents, and where it runs.
Two AWS platforms run agents. Amazon Bedrock Agents is a managed, configuration-based agent service. Since July 30, 2026, AWS calls it Amazon Bedrock Agents Classic: it is in maintenance mode, closed to new customers, and will not receive new features, though existing agents and all its APIs will keep working. Amazon Bedrock AgentCore is the platform AWS now recommends for new agents. It runs agents built with any framework and any model. Exam scenarios and existing workloads still use Bedrock Agents terms, so this lesson teaches both.
An agent deployment is done when each tool is reachable and authorized, the agent answers through a stable endpoint, and a test request shows the expected tool calls.
The diagram below shows the parts of an agent and the two protocols that connect it to the outside world.
Agent building blocks: Model, instructions, and tools
Every agent, on either platform, is assembled from the same parts:
Foundation model: The model does the reasoning and chooses tools. It must support tool use.
Instructions (system prompt): They state the agent’s job and limits. “Handle order questions; never issue refunds above $500” is better than “be a helpful assistant.”
Tools: Each tool has a name, a natural-language description, and a typed input schema. The model chooses a tool by reading its description, so a vague description causes wrong or missing tool calls. ...