AWS MLA-C01 MLA-C02: What Changed (Spoiler: Bedrock Is In)

AWS is revising the Machine Learning Engineer – Associate exam from MLA-C01 to MLA-C02. I went through the official comparison page and both exam guides side by side to lay out exactly what changed — and what it means if you're preparing for it.

The short version: generative AI and AI agents — in other words, Amazon Bedrock — are now in scope.

Update: my MLA course has been updated for the C02 changes.

📚 AWS Certified Machine Learning Engineer – Associate course on Udemy:
https://www.udemy.com/course/aws-machine-learning-engineer-associate-mla-en

🎫 Coupons

Free (limited seats): https://www.udemy.com/course/aws-machine-learning-engineer-associate-mla-en/?couponCode=FREESEP
Discount (if the free seats are gone): https://www.udemy.com/course/aws-machine-learning-engineer-associate-mla-en/?couponCode=DISCOUNTSEP
Both expired? I issue new coupons every month on YouTube and Udemy → https://www.youtube.com/@MaruchinTech-cloud/posts

1. Positioning and candidate profile

The C02 description was rewritten around "AI and ML solutions," and now explicitly states that the exam validates the ability to work with both traditional ML models and foundation models (FMs).

Item	MLA-C01	MLA-C02
Candidate experience	1+ year with SageMaker AI	1+ year with SageMaker AI plus 1+ year with Amazon Bedrock — experience in both traditional ML and generative AI
Recommended knowledge (added)	—	Understanding FM capabilities, limitations, and use cases; Bedrock features
Recommended knowledge (removed)	Experience with code repositories for version control	—
Out-of-scope tasks (removed)	Analyzing model quantization and its effect on accuracy	— (consistent with the removal of model-compression skills below)

2. Exam format

Item	MLA-C01	MLA-C02
Questions	65 (50 scored + 15 unscored)	Same
Passing score	720 / 1000	Same (no pass/fail during beta)
Question types	Multiple choice, multiple response, ordering, matching	Multiple choice and multiple response only

Ordering and matching — the new formats introduced with C01 — are gone from the guide.

3. Domain weighting

MLA-C01	MLA-C02
D1 Data Preparation for ML (28%)	D1 Data Preparation for ML and AI (28%)
D2 ML Model Development (26%)	D2 ML Model and Foundation Model (FM) Development (24%)
D3 Deployment and Orchestration of ML Workflows (22%)	D3 Deployment and Orchestration of ML and AI Workflows (24%)
D4 ML Solution Monitoring, Maintenance, and Security (24%)	D4 Operating, Monitoring, and Securing ML and AI Solutions (24%)

Two points moved from D2 to D3. The four-domain, twelve-task skeleton is otherwise preserved 1:1 — the official comparison page maps every task.

4. Skills added

Everything the official page lists as new, grouped by domain as keywords:

Domain	Added skills
D1 Data preparation	Vector database configuration (OpenSearch Service, RDS + pgvector, S3), ingesting and storing multimodal data (text, image, audio), configuring and using embedding models, advanced text preprocessing (tokenization, domain-specific augmentation), document preparation for RAG (chunking strategies, metadata extraction), data masking / redaction / anonymization, data prep for FM fine-tuning, continued pre-training, and distillation, bias metrics on multimodal data, training-data integrity validation (prompt-response pair validation, content safety screening), data cleansing (outlier detection, missing-value imputation, deduplication)
D2 Model / FM development	FM selection from Bedrock, FM fine-tuning strategies, tradeoffs between custom / managed / pre-trained / FM approaches, choosing RAG architecture patterns, performance-latency-cost tradeoffs, customization via prompt engineering and fine-tuning, optimizing retrieval components and embedding models, human evaluation frameworks (HITL), NLP metrics (BLEU, ROUGE, BERTScore, semantic similarity), AI evaluation (output evaluation, bias detection, LLM-as-a-judge), RAG system monitoring (retrieval accuracy)
D3 Deployment and orchestration	Evaluating FM deployment options, bringing in models built outside AWS (SageMaker AI, Bedrock Custom Model Import), deploying and configuring agents (service and tool integration, agent communication protocols), FM hosting and resource allocation, RAG configuration (retrieval strategies, reranking), creating and managing Bedrock Knowledge Bases, agent state management, scaling for GPU workloads, deploying agentic workflow infrastructure, Bedrock Prompt Management, automated agent deployment and versioning, AI model test frameworks including prompt tests, versioned deployment of fine-tuned models, pipelines for RAG updates and knowledge-base re-ingestion
D4 Operations, monitoring, security	Monitoring agent performance and coordination (coordination failures, streaming disconnects, tool failures), FM performance monitoring (Bedrock evaluations), FM cost evaluation for production inference, agent resource consumption, AI-specific cost patterns (token usage, embedding compute, vector DB storage), code and image vulnerability scanning in CI/CD (CodeGuru, Inspector), credential types for FM access (Bedrock API keys, IAM), Bedrock Guardrails for safeguards and sensitive-data protection

Almost every addition sits in the Bedrock-centered generative AI, RAG, and agent space.

5. Skills removed

Task	Removed
1.3	Configuring data loading to training resources (EFS, FSx)
2.2	Fine-tuning pre-trained models on custom datasets (Bedrock, JumpStart)
2.2	Reducing model size (data type changes, pruning, compression)
3.1	Optimizing for edge devices (SageMaker Neo)
3.2	Bring your own container (BYOC) on SageMaker
4.2	Infrastructure monitoring with EventBridge
4.2	Troubleshooting capacity issues (provisioned concurrency, service quotas, auto scaling)

Fine-tuning didn't disappear — it was redefined in the FM context (C02 tasks 1.2.9 and 2.2.8). What got lighter is the traditional-ML periphery: SageMaker Neo, BYOC, model compression.

6. In-scope services

Change	Services
Added in C02	Amazon Bedrock AgentCore, AWS CodeConnections, Amazon Inspector
Renamed	Amazon SageMaker → Amazon SageMaker AI
Removed in C02	Amazon A2I, Amazon Fraud Detector, Amazon Kendra, Lookout for Equipment / Metrics / Vision, Amazon Mechanical Turk, Amazon Q, AWS Chatbot, AWS Serverless Application Repository, CloudWatch Logs (now folded into CloudWatch)

Amazon Q and Kendra dropping out reflects RAG's center of gravity moving to Bedrock Knowledge Bases.

7. An instructor's read

Until now the split among AWS AI certifications was clear: MLA-C01 was SageMaker-centered, AIP-C01 was Bedrock-centered. With C02, Bedrock enters MLA too.

AIP-C01 tests Bedrock at a high difficulty level, so my expectation is that MLA-C02 tests it at up to a moderate level. If you passed C01, or are studying for it now, the new items to cover are:

Bedrock Knowledge Bases, Prompt Management, Guardrails, Custom Model Import, Model Evaluation
Bedrock AgentCore (including Observability), CloudWatch generative AI observability
Vector search with pgvector / OpenSearch Service
NLP evaluation metrics (BLEU / ROUGE / BERTScore) and LLM-as-a-judge

The backbone of your preparation doesn't change much. SageMaker-centered traditional ML plus the core Bedrock features should cover it. I'll update this article if the beta or GA version changes anything.

Wrapping up

MLA-C02 is a "traditional ML + foundation models" exam; 1+ year of Bedrock experience is now part of the candidate profile
The 4-domain, 12-task structure is preserved; only 2 points moved from D2 to D3
Added skills cluster around Bedrock-centered generative AI, RAG, and agents
Removed skills are traditional-ML periphery: SageMaker Neo, BYOC, model compression
Ordering and matching question types are gone; multiple choice and multiple response only