Cloud concept · Updated

AI Agents - AI agents are software systems that use artificial intelligence to select and carry out actions toward a goal within defined permissions and constraints.

In short

In foundation model applications, an AI agent combines a model with instructions, context, and tools to complete tasks that may require several decisions and interactions with external systems. Agents matter because they can move beyond generating answers to performing business work, but their access, reliability, cost, and degree of autonomy require explicit controls.

Cloud Concepts Hub →

6 key aspects of AI Agents

AI agents combine goal-directed decisions, tool execution, and task context with explicit limits, oversight, and operational evaluation.

01
Goal-directed action

An agent selects actions that advance a task rather than only producing text. Define measurable completion criteria, such as a confirmed appointment or a successfully created support ticket.

02
Tool use and external integration

Tools connect the agent to APIs, databases, and business applications. Validate every request and enforce permissions in execution code rather than trusting model-generated arguments.

03
Iterative execution and state

The agent uses tool results and task context to choose subsequent actions. Maintain enough state to resume safely, and apply iteration limits and stopping conditions to prevent uncontrolled execution.

04
Bounded autonomy and human oversight

Grant only the actions needed for the task and require approval for sensitive or irreversible changes. Enforce approval gates outside the model so they cannot be bypassed by generated output.

05
Grounding and context management

Relevant documents, conversation history, and current system data help the agent make informed decisions. Keep context access-controlled and treat retrieved content as untrusted input that may contain misleading instructions.

06
Operational evaluation and reliability

Measure successful task completion, unauthorized-action attempts, latency, and total execution cost. Test dependency failures, ambiguous requests, and repeated tool calls before enabling production writes.

Cloud services for AI Agents

Amazon Bedrock can supply foundation model inference and managed agent orchestration, while AWS Lambda exposes controlled business operations as tools. AWS Step Functions can surround those operations with deterministic sequencing, retry policies, and approval waits when stronger workflow control is needed. IAM limits the AWS resources each component can access, while application logic checks the requesting user's authority and business rules. CloudWatch provides operational visibility through instrumented logs, metrics, and alarms. Together, these services support a design in which the model proposes actions, execution components enforce constraints, and operators can investigate outcomes without treating model output as proof that an action succeeded.

Amazon Bedrock

Amazon Bedrock provides access to foundation models, and Agents for Amazon Bedrock provides managed capabilities for coordinating model-driven tasks with knowledge sources and action groups. Applications still need appropriate tool implementations, permissions, and business controls.

AWS LambdaAWS Lambda

AWS Lambda can implement tools that an agent invokes, such as checking order status or creating a ticket. Functions provide an execution boundary for input validation, authorization checks, and controlled access to downstream systems.

AWS Step FunctionsAWS Step Functions

AWS Step Functions can coordinate controlled workflows around agent decisions, including retries, timeouts, and approval waits through callback patterns. It is useful when model-driven choices must fit within an explicit, auditable execution sequence.

Amazon CloudWatchAmazon CloudWatch

Amazon CloudWatch can collect execution logs and operational metrics from agent components and tools. Instrument the application to detect failures, excessive iterations, and latency, while redacting sensitive content from logs.

AI Agents exam questions

Real practice exam questions on AI Agents, with the correct answer and its explanation. Answer them with every option in the exam simulator.

AIB-C01 (beta)

A maintenance company uses an AI agent to investigate equipment faults, select an appropriate replacement part, and request a repair appointment through approved tools. The agent must retain flexibility to choose diagnostic checks. After a part is reserved, however, company policy requires the reservation to be released if appointment scheduling fails. If releasing the reservation also fails, the case must be referred to an employee. The company wants centrally tracked execution and minimal custom coordination code, without relying on the model to remember cleanup actions. Which solution best meets these requirements?

Show answerCORRECT ANSWER

Use AWS Step Functions to coordinate the agent and booking tools, with explicit failure paths that release reserved parts and route unsuccessful releases to an employee.

Step Functions can preserve the agent's flexible diagnostic behavior while enforcing predefined recovery paths around business actions. Explicit cleanup and escalation steps provide centrally tracked orchestration without depending on the model to decide whether cleanup is necessary.

MLA-C02 (beta)

A company operates a multi-turn troubleshooting agent that invokes an Amazon Bedrock foundation model through the Converse API. Each model request starts with a long, shared policy prefix configured for prompt caching. The selected model supports prompt caching and returns cache-usage metadata. The agent also calls diagnostic tools whose results are appended after the shared prefix.

The company wants to determine how effectively the prompt cache is reused throughout the day for each policy-template version. Engineers must measure actual cache reuse, rather than cache eligibility or reuse of complete application responses. Policy text, user messages, and tool results must not be logged.

Which monitoring approach best meets these requirements?

Show answerCORRECT ANSWER

Record inputTokens, outputTokens, cacheReadInputTokens, and cacheWriteInputTokens from each Converse response's usage metadata, together with the policy-template version and timestamp. Aggregate cache-read and cache-write token volumes by version and hour in CloudWatch without recording request or response content.

The response usage metadata reports actual cache reads and writes for supported models. Aggregating these measurements alongside ordinary token usage reveals when each policy-template version benefits from cache reuse and when requests populate the cache instead. Only usage metadata and operational identifiers need to be recorded.

AIF-C01

A software company is evaluating an AI assistant that creates project workspaces by selecting and calling tools to create a repository, assign team permissions, and register the project. Identical requests sometimes produce different outcomes. The company wants to estimate how reliably a user request will succeed during one execution under the intended production settings, without a user restarting the task. Which evaluation plan BEST meets this objective?

Show answerCORRECT ANSWER

Run each representative request multiple times from an equivalent clean test state using production settings, and measure the percentage of executions that produce the required workspace and permissions.

Repeated independent executions capture variability in the agent's decisions and tool use. Resetting the test state prevents previous executions from influencing later results. Verifying actual workspace state measures successful task completion rather than a plausible completion message.

Practice AI Agents in a full exam

Questions on AI Agents in each practice exam, with every answer option and an explanation for each:

Start a free practice exam

What is AI Agents

2 min read

An AI agent is a software system that selects actions to achieve a goal, observes the results, and decides what to do next. In foundation model applications, the model commonly interprets the request, proposes intermediate steps, and selects tools. An orchestration layer manages model calls, tool execution, state, and stopping conditions. The model does not inherently have access to business systems: the application supplies interfaces and enforces permissions.

A typical interaction follows a bounded action-and-observation cycle. A user asks a support agent to investigate a delayed order. The agent requests missing order details, retrieves shipment information through an approved API, checks the applicable policy, and offers an authorized resolution. If a refund exceeds a configured threshold, the workflow requests human approval instead of allowing the model to decide whether approval is necessary. The agent then reports the actual outcome rather than merely generating a plausible completion message.

Cloud implementations usually combine a hosted model, an agent runtime or orchestration component, API-backed tools, and storage for task state. Tools might query inventory, create support tickets, schedule appointments, or retrieve documents. Tool definitions describe supported operations and their inputs. Execution code must independently validate arguments, authenticate the caller, and check authorization. Retrieved documents provide evidence, but their contents should be treated as untrusted data rather than instructions that can override application policy.

Business applications include customer service agents that resolve account issues, IT operations agents that collect diagnostics and propose remediation, procurement agents that compare approved suppliers, and sales agents that update customer records after a conversation. Their value is greatest when work requires interpreting variable inputs and choosing among permitted actions. A fixed workflow is often preferable when the sequence is predictable, strict reproducibility is required, or errors have unacceptable consequences. Agents and deterministic workflows can also be combined, with flexible interpretation inside tightly controlled execution paths.

Agent autonomy creates trade-offs. Multiple model calls and tool requests increase latency and cost, and an incorrect early decision can affect later steps. Engineers should impose iteration limits, timeouts, spending limits, and clear escalation conditions. Retries need special care: repeating a read operation is usually less risky than repeating a payment or refund. Idempotency keys, transaction controls, and durable task state help prevent duplicate actions. Evaluate completed business outcomes, tool-selection accuracy, policy compliance, and behavior when dependencies fail, rather than judging only the final response.

Several misconceptions are important. A chatbot is not necessarily an agent; answering questions alone does not demonstrate goal-directed action. Retrieval-augmented generation can supply relevant information without providing an action loop. Agents do not necessarily learn from each interaction or require model fine-tuning, and they need not operate without supervision. Effective designs use least-privilege access, minimize sensitive data exposure, record execution events, and require approval for consequential actions. Safety instructions and output filters complement these controls but do not replace authorization or business-rule enforcement.

Learn AI Agents in lessons

6 short lessons on AI Agents in 3 certification courses.

Open the lessons