Hands-on Lab — AIP-C01 AWS Certified Generative AI Developer - Professional

Overview

Prerequisites

Steps

1. 1.Pick our Terraform version and AWS region

   Standard opener. us-east-1 for broadest Bedrock + Bedrock Agent feature availability. The provider version pin is non-negotiable: aws_bedrockagent_* resources require ~> 5.60 or later.

   Copy

   terraform {
     required_version = ">= 1.5"
   
     required_providers {
       aws = {
         source  = "hashicorp/aws"
         version = "~> 5.60"
       }
       archive = {
         source  = "hashicorp/archive"
         version = "~> 2.4"
       }
     }
   }
   
   provider "aws" {
     region = "us-east-1"
   
     default_tags {
       tags = {
         Project   = "certlabpro-aip-c01"
         ManagedBy = "terraform"
       }
     }
   }
   
   data "aws_caller_identity" "current" {}
   data "aws_region" "current" {}

2. 2.Give the agent a private bucket for any reference data it might read

   Provisions:

   • Amazon S3

   Most production Bedrock Agents end up reading something from S3 — knowledge-base source documents, JSON tool definitions, response templates. We provision a bucket up front so we don't have to plumb permissions in later. Encryption + public-access-block defaults are the AIP-C01 Foundation Model Integration, Data Management, and Compliance domain at work.

   The AIP-C01 exam tests one specific compliance pattern around this bucket: client-side encryption with a key the application controls is the right answer when the regulation says "AWS must never have the unencrypted data". For everything else, SSE-S3 (what we do here) is sufficient.

   Copy

   resource "aws_s3_bucket" "agent_data" {
     bucket_prefix = "certlabpro-aip-c01-"
   }
   
   resource "aws_s3_bucket_public_access_block" "agent_data" {
     bucket = aws_s3_bucket.agent_data.id
   
     block_public_acls       = true
     block_public_policy     = true
     ignore_public_acls      = true
     restrict_public_buckets = true
   }
   
   resource "aws_s3_bucket_server_side_encryption_configuration" "agent_data" {
     bucket = aws_s3_bucket.agent_data.id
   
     rule {
       apply_server_side_encryption_by_default {
         sse_algorithm = "AES256"
       }
     }
   }

3. 3.Add a production-grade guardrail with content + PII + denied-topics filters

   Provisions:

   • Amazon Bedrock Guardrails

   AIP-C01's AI Safety, Security, and Governance domain (20% of the exam) goes much deeper than the AIF-C01 version. The exam tests the multi-layered guardrail composition: content filters for harmful categories at high strength, sensitive-information policies for PII (with BLOCK rather than the AIF-style ANONYMIZE — production agents usually need hard blocks), and topic policies for application-specific forbidden topics.

   We configure all three. The denied topic "FinancialAdvice" in the policy is a common production guardrail for non-financial chatbots — even Claude 3 will sometimes drift into giving stock tips unless you tell the guardrail to block topical drift. AIP-C01 tests this exact composition pattern in scenario questions where the prompt is benign but the response could go off-policy.

   Copy

   resource "aws_bedrock_guardrail" "production" {
     name                      = "certlabpro-aip-c01-production"
     description               = "Production-grade safety rail for the AIP-C01 lab agent."
     blocked_input_messaging   = "I can't help with that request."
     blocked_outputs_messaging = "I can't share that response."
   
     content_policy_config {
       filters_config {
         input_strength  = "HIGH"
         output_strength = "HIGH"
         type            = "HATE"
       }
       filters_config {
         input_strength  = "HIGH"
         output_strength = "HIGH"
         type            = "VIOLENCE"
       }
       filters_config {
         input_strength  = "HIGH"
         output_strength = "HIGH"
         type            = "SEXUAL"
       }
       filters_config {
         input_strength  = "HIGH"
         output_strength = "HIGH"
         type            = "INSULTS"
       }
       filters_config {
         input_strength  = "HIGH"
         output_strength = "HIGH"
         type            = "MISCONDUCT"
       }
     }
   
     sensitive_information_policy_config {
       pii_entities_config {
         action = "BLOCK"
         type   = "EMAIL"
       }
       pii_entities_config {
         action = "BLOCK"
         type   = "PHONE"
       }
       pii_entities_config {
         action = "BLOCK"
         type   = "US_SOCIAL_SECURITY_NUMBER"
       }
       pii_entities_config {
         action = "BLOCK"
         type   = "CREDIT_DEBIT_CARD_NUMBER"
       }
     }
   
     topic_policy_config {
       topics_config {
         name       = "FinancialAdvice"
         definition = "Specific recommendations to buy, sell, or hold financial instruments."
         examples   = ["Should I buy NVDA?", "Is now a good time to sell my Bitcoin?"]
         type       = "DENY"
       }
     }
   }

4. 4.Create the Bedrock Agent with a scoped IAM role

   Provisions:

   • Amazon Bedrock
   • AWS IAM

   The agent ties together a foundation model, a system prompt, and (in Step 5) one or more action groups that extend its capabilities. The IAM role we attach is the agent's execution role — it's what the agent uses to invoke models, retrieve from knowledge bases, and call action-group Lambdas. AIP-C01's Implementation and Integration domain tests this exact role/trust shape: bedrock.amazonaws.com as the principal, scoped permissions to bedrock:InvokeModel on specific model ARNs, and lambda:InvokeFunction on specific function ARNs.

   The agent's instruction field is the system prompt — short, action-oriented, naming the agent's purpose. AIP-C01 questions about prompt engineering at agent level test exactly this: the instruction is the persistent personality; the user message changes per turn.

   Copy

   resource "aws_iam_role" "agent" {
     name = "certlabpro-aip-c01-agent"
   
     assume_role_policy = jsonencode({
       Version = "2012-10-17"
       Statement = [{
         Effect    = "Allow"
         Principal = { Service = "bedrock.amazonaws.com" }
         Action    = "sts:AssumeRole"
         Condition = {
           StringEquals = {
             "aws:SourceAccount" = data.aws_caller_identity.current.account_id
           }
         }
       }]
     })
   }
   
   resource "aws_iam_role_policy" "agent_invoke_model" {
     name = "invoke-foundation-model"
     role = aws_iam_role.agent.id
   
     policy = jsonencode({
       Version = "2012-10-17"
       Statement = [{
         Effect = "Allow"
         Action = ["bedrock:InvokeModel", "bedrock:InvokeModelWithResponseStream"]
         Resource = "arn:aws:bedrock:${data.aws_region.current.name}::foundation-model/anthropic.claude-3-haiku-20240307-v1:0"
       }]
     })
   }
   
   resource "aws_bedrockagent_agent" "main" {
     agent_name                  = "certlabpro-aip-c01-agent"
     agent_resource_role_arn     = aws_iam_role.agent.arn
     foundation_model            = "anthropic.claude-3-haiku-20240307-v1:0"
     instruction                 = "You are a helpful assistant for the certlabpro AIP-C01 lab. Answer concisely. When the user asks about the current time, use the get_current_time tool."
     idle_session_ttl_in_seconds = 600
   
     # Attach the guardrail from Step 3. Bedrock evaluates inputs and outputs
     # against the guardrail policy on every model call.
     guardrail_configuration {
       guardrail_identifier = aws_bedrock_guardrail.production.guardrail_id
       guardrail_version    = aws_bedrock_guardrail.production.version
     }
   }

5. 5.Add an Action Group that lets the agent invoke a Lambda tool

   Provisions:

   • Amazon Bedrock
   • AWS Lambda
   • AWS IAM

   Action Groups are how Bedrock Agents call out to your code. The pattern: define the API schema (what the tool accepts and returns), point at a Lambda that implements the actions, and Bedrock handles the rest — when the model decides it needs the tool, it calls your Lambda with structured arguments and feeds the response back into the conversation.

   We define a minimal get_current_time action with the OpenAPI 3.0 schema Bedrock expects. The Lambda implementation is intentionally tiny — what the exam tests is the shape of the integration (function-calling Lambda + OpenAPI schema + Bedrock invocation permission), not the function's internal logic.

   The aws_lambda_permission resource is the canonical AIP-C01 gotcha: without it, the agent's invocation of the Lambda gets AccessDenied even though every other piece is wired correctly. Bedrock Agent → Lambda is an external-service invocation pattern; the function's resource policy needs to allow bedrock.amazonaws.com explicitly.

   With Action Groups in place, the full agent shape is complete: model + system prompt + safety guardrail + tool-calling extensibility, all under one scoped IAM identity. Every additional AIP-C01 pattern (Knowledge Base retrieval, multi-action agents, agent-collaboration via supervisor agents) extends this base.

   Copy

   # ── Lambda tool ───────────────────────────────────────────────
   data "archive_file" "tool" {
     type        = "zip"
     output_path = "${path.module}/build/tool.zip"
   
     source {
       filename = "index.py"
       content  = <<-EOT
         import json
         from datetime import datetime, timezone
         def handler(event, context):
           # Bedrock Agent passes the parsed args under `requestBody` / `parameters`.
           # For a simple no-arg tool, just return the current time.
           now_iso = datetime.now(timezone.utc).isoformat()
           return {
             "messageVersion": "1.0",
             "response": {
               "actionGroup": event.get("actionGroup", ""),
               "apiPath": event.get("apiPath", "/"),
               "httpMethod": event.get("httpMethod", "GET"),
               "httpStatusCode": 200,
               "responseBody": {"application/json": {"body": json.dumps({"current_time_utc": now_iso})}},
             },
           }
       EOT
     }
   }
   
   resource "aws_iam_role" "tool_lambda" {
     name = "certlabpro-aip-c01-tool"
   
     assume_role_policy = jsonencode({
       Version = "2012-10-17"
       Statement = [{
         Effect    = "Allow"
         Principal = { Service = "lambda.amazonaws.com" }
         Action    = "sts:AssumeRole"
       }]
     })
   }
   
   resource "aws_iam_role_policy_attachment" "tool_logs" {
     role       = aws_iam_role.tool_lambda.name
     policy_arn = "arn:aws:iam::aws:policy/service-role/AWSLambdaBasicExecutionRole"
   }
   
   resource "aws_lambda_function" "tool" {
     function_name    = "certlabpro-aip-c01-tool"
     role             = aws_iam_role.tool_lambda.arn
     runtime          = "python3.12"
     handler          = "index.handler"
     filename         = data.archive_file.tool.output_path
     source_code_hash = data.archive_file.tool.output_base64sha256
     timeout          = 10
   }
   
   resource "aws_lambda_permission" "agent_invoke_tool" {
     statement_id  = "AllowBedrockAgentInvoke"
     action        = "lambda:InvokeFunction"
     function_name = aws_lambda_function.tool.function_name
     principal     = "bedrock.amazonaws.com"
     source_arn    = aws_bedrockagent_agent.main.agent_arn
   }
   
   # Also let the agent invoke the lambda from its execution role.
   resource "aws_iam_role_policy" "agent_invoke_lambda" {
     name = "invoke-tool"
     role = aws_iam_role.agent.id
   
     policy = jsonencode({
       Version = "2012-10-17"
       Statement = [{
         Effect   = "Allow"
         Action   = "lambda:InvokeFunction"
         Resource = aws_lambda_function.tool.arn
       }]
     })
   }
   
   # ── Action Group ──────────────────────────────────────────────
   resource "aws_bedrockagent_agent_action_group" "time" {
     action_group_name          = "time-tools"
     agent_id                   = aws_bedrockagent_agent.main.agent_id
     agent_version              = "DRAFT"
     description                = "Time-related tool calls for the agent."
     skip_resource_in_use_check = false
   
     action_group_executor {
       lambda = aws_lambda_function.tool.arn
     }
   
     api_schema {
       payload = jsonencode({
         openapi = "3.0.0"
         info    = { title = "Time tools", version = "1.0.0" }
         paths = {
           "/get_current_time" = {
             get = {
               operationId = "get_current_time"
               description = "Returns the current UTC time as an ISO-8601 string."
               responses = {
                 "200" = {
                   description = "Current UTC time"
                   content = {
                     "application/json" = {
                       schema = {
                         type = "object"
                         properties = {
                           current_time_utc = { type = "string", format = "date-time" }
                         }
                       }
                     }
                   }
                 }
               }
             }
           }
         }
       })
     }
   }

Cleanup