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.
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) |
| 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.
| 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.
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.
| Task | Removed |
|---|---|
| 1.3 | Configuring data 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.
| 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.
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:
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.
About the author
Maruchin Tech — 12x AWS Certified | Cloud & AI for manufacturing and supply chain (AWS / Google Cloud / Azure) | Udemy instructor (100K+ students)
📚 My AI certification courses (AIF / MLA / AIP):
[https://www.udemy.com/user/maruchin-tech-2/](https://www.udemy.com/user/maruchin-tech-2/)
👨🏫 All my courses:
(Eng) [https://www.udemy.com/user/maruchin-tech-2/](https://www.udemy.com/user/maruchin-tech-2/)
(Jpn) [https://www.udemy.com/user/shan-wang-wan-jun-2/](https://www.udemy.com/user/shan-wang-wan-jun-2/)
🎫 Monthly discount coupons:
https://www.youtube.com/@MaruchinTech-cloud/posts