Domain 1 introduced agentic AI as AI that plans and acts toward a goal. This lesson opens up the parts that make an agent work: tools, memory and orchestration. It then covers how several agents can work as a team, and how the Model Context Protocol (MCP) gives agents a standard way to connect to outside systems.
Tools
A tool is any capability an agent can call to act on the world or fetch information: an API, a database query, a search, a calculator or a function in your own code. The model doesn't run the tool itself. It decides that a tool is needed and which inputs to pass. The application runs the tool and hands the result back, and the model uses that result to continue. In Amazon Bedrock this is called tool use. Tools are what let an agent check stock, book a meeting or file a ticket, rather than only describe how.
Memory
Models are stateless, so agents need memory added around them. Short-term memory tracks what has happened within the current task or session, such as earlier turns and steps completed so far. Long-term memory keeps useful facts across sessions, such as a customer's preferences or the outcome of past cases. Amazon Bedrock AgentCore Memory provides both as a managed service, and multi-agent systems can also share memory so agents stay in sync.
Workflow orchestration
Orchestration is the logic that runs the agent's loop: plan a step, call a tool, look at the result, decide the next step, and stop when the goal is met or a limit is reached. Some workflows are tightly structured, with a coordinator following a set sequence of steps, which suits predictable business processes. Others give the model more freedom to choose its own path, which suits open-ended problems but needs guardrails and limits on what the agent is allowed to do.
Multi-agent patterns
Complex jobs can be split across several specialised agents instead of one agent doing everything. Two broad patterns appear in AWS guidance:
| Pattern | How it works | Good for |
|---|---|---|
| Supervisor (coordinator) | One agent receives the request, breaks it up and routes parts to specialist collaborator agents | Structured business processes, such as a mortgage assistant with agents for new loans, existing loans and general questions |
| Peer collaboration | Agents with distinct roles, such as planner, researcher and critic, share information and refine each other's work | Open-ended research, creative or exploratory tasks |
Agents communicate through messages, shared memory or by passing results along a chain. Specialised agents are easier to build, test and secure than one agent with every tool and permission.
Model Context Protocol
The Model Context Protocol (MCP) is an open protocol that standardises how AI applications connect to external tools and data. An MCP server exposes capabilities, such as tools, and an MCP client inside an AI application can discover those tools, call them with parameters and receive the results. Because the interface is standard, a tool wrapped once as an MCP server can be used by any MCP-compatible agent, instead of building a custom integration for every pairing. AWS describes it as a universal connector. Amazon Bedrock AgentCore Gateway, for example, converts existing APIs and AWS Lambda functions into MCP-compatible tools. A separate open protocol, Agent2Agent (A2A), focuses on agents communicating with other agents.
On the exam
Match the scenario to the part. "The agent needs to check live inventory" is a tool. "It should remember the customer next week" is long-term memory. "Many teams' tools must work with many agents without custom integrations" is MCP. "Route each request to a specialist agent" is the supervisor pattern. Note that Amazon Bedrock Agents is now called Bedrock Agents Classic and is closed to new customers; AWS points new builds to Amazon Bedrock AgentCore.