An AI agent uses a foundation model to plan and act toward a goal: it calls tools, remembers context and decides its next step. These sections cover single agents, the Model Context Protocol and multi-agent systems.
Agents, tools and MCP
An agent is a foundation model given a goal, tools and memory, running in a loop: think, act, observe, repeat. The Model Context Protocol (MCP) standardizes how agents reach tools and data.
How it works
- The agent receives a goal and its context (instructions, memory, available tools).
- The model plans the next step and decides whether to call a tool.
- The tool runs (search, database, API, code) and returns a result.
- The agent repeats until the goal is met or a human approval is needed.
| Need | Build |
|---|---|
| One-shot task (summarize, classify) | Direct model call |
| Fixed steps in a known order | Workflow with model calls inside steps |
| Reasoning plus tools over several steps | Single agent |
| Complex work across specialties | Multiple agents |
Exam cues
- If the question says connect an agent to many tools and data sources in a standard way, think MCP.
- If the question says the system decides its own steps and calls tools, think AI agent.
- If the question says steps are fixed and known, think a predefined workflow.
- If the question says a person must approve risky actions, think human-in-the-loop.
Remember
- MCP: open standard; resources, tools, prompts.
- Use the least agentic design that solves the task.
Trap: More autonomy isn’t always better. Agents are less predictable than workflows; use them where reasoning and tool choice are needed.
From the AWS docs (8)
- An AI agent is an autonomous software system that uses an LLM as its reasoning engine to perceive context, plan actions, execute tasks and adapt toward a defined goal, taking multi-step actions and invoking external tools.
- A tool is a function or API an agent can invoke to interact with external systems, such as web search, database queries, code execution or API calls.
- A workflow is a structured sequence of agent actions, tool calls and decision points, implemented for example as an AWS Step Functions state machine, an agent reasoning loop or both.
- Tasks a model completes in one request, such as summarization or classification, need only a direct model call; tasks that need reasoning, tool use or several steps call for an agent; complex workflows can use multiple specialized agents.
- The Model Context Protocol (MCP) is an open protocol, introduced by Anthropic, that connects LLM applications to external data sources and tools through one standard interface instead of a custom integration per system.
- MCP exposes three kinds of capability: resources (data and content such as files, databases and APIs), tools (actions such as creating tickets or running queries) and prompts (reusable prompt templates).
- Human-in-the-loop brings human judgment into agent workflows at defined decision points, from notification only to approval required before the agent proceeds.
- Agentic AI use cases include supporting research and development, code transformation, incident response automation and customer service automation.
Multi-agent systems and memory
When one agent can’t hold all the skills or context, a multi-agent system splits the work among specialized agents, coordinated by an orchestrator or by peers. Memory lets agents carry context within a session and across sessions.
How it works
| Pattern | How agents coordinate |
|---|---|
| Supervisor-worker | An orchestrator delegates subtasks and combines results |
| Pipeline | Each agent’s output feeds the next |
| Swarm | Agents work in parallel and share a workspace |
| Peer-to-peer | Agents talk directly with each other |
| Dynamic graph | Routing between agents is decided at run time |
- Short-term memory keeps the current conversation; long-term memory keeps knowledge across sessions.
- MCP connects agents to tools; A2A connects agents to other agents.
- A workflow engine such as AWS Step Functions can orchestrate fixed parts of the flow.
Exam cues
- If the question says one agent breaks down requests and delegates to specialists, think supervisor (orchestrator) pattern.
- If the question says remember within a session versus across sessions, think short-term versus long-term memory.
- If the question says agents communicating with agents, think A2A; agents to tools, think MCP.
Remember
- Specialized agents do one thing well; orchestrators coordinate.
- Memory tiers: short-term (session) and long-term (persistent).
Trap: MCP and A2A solve different problems. MCP is agent-to-tool; A2A is agent-to-agent.
From the AWS docs (7)
- A multi-agent system has specialized agents collaborate on tasks beyond one agent, using patterns such as supervisor-worker (hierarchical), peer-to-peer, swarm, pipeline, dynamic graph or hybrid.
- An orchestrator agent breaks a complex task into subtasks, delegates them to specialized agents, coordinates the information between them and combines the results.
- A specialized agent performs a single, well-defined capability and is the building block of a multi-agent system.
- In a swarm, agents work independently on overlapping parts of a problem and share findings through a common workspace, without central supervision.
- Short-term (session) memory holds context for the current task or conversation; long-term (persistent) memory keeps preferences, past outcomes and domain knowledge across sessions.
- Agent orchestration can be done by an orchestrator agent, a workflow engine such as AWS Step Functions, or both, from static workflows defined at design time to dynamic graphs.
- Amazon Bedrock AgentCore supports agent protocols such as MCP and Agent2Agent (A2A).
Flashcards
Nothing to review here right now. Pick other topics or another filter.
·
All 12 cards as a list
What makes an AI agent different from a single call to a foundation model?
An agent plans multi-step actions toward a goal, calls tools (APIs, databases, code) and keeps state, instead of answering one prompt once.
An AI assistant must search files, query a database and open tickets in other systems. What gives it a standard way to connect to all of them?
The Model Context Protocol (MCP), an open standard for connecting agents to tools and data sources.
Name the three kinds of capability an MCP server can expose.
Resources (data), tools (actions) and prompts (reusable templates).
A process always runs the same steps in a fixed order, with a model call inside two steps. Does it need an autonomous agent?
No. A predefined workflow (for example AWS Step Functions) with model calls inside steps is simpler and more predictable.
An agent can issue refunds. How do you keep a person in control of large refunds?
Add a human-in-the-loop approval step, so the agent can't proceed past that decision point without explicit human approval.
Name three good uses for agentic AI.
Any three of: customer service automation, incident response automation, code transformation, research support, multi-step workflow automation.
A company wants to classify support emails by topic. Does it need an AI agent?
No. One model call (or a classifier) does it; agents fit tasks needing reasoning, tools and several steps.
A support request needs billing, shipping and technical knowledge. One agent splits the request and sends parts to three expert agents, then combines their answers. Which pattern is this?
Supervisor (orchestrator) multi-agent pattern with specialized agents.
A chatbot must remember earlier turns in today's chat and a customer's preferences from past months. Which memory types cover each?
Short-term (session) memory for today's conversation; long-term (persistent) memory for preferences across sessions.
Several research agents explore a problem in parallel and share findings in a common workspace with no central coordinator. Which pattern is this?
A swarm.
Which protocol lets agents built with different frameworks communicate with each other, as supported by Amazon Bedrock AgentCore?
Agent2Agent (A2A). MCP connects agents to tools and data.
Is a multi-agent system always better than one agent with several tools?
No. Multiple agents add coordination cost and failure points; use them when the task needs separate specialties.