Azure AI Apps and Agents Developer (AI-103) Study Plan for 2026

Topics

• • Python SDK calls, JSON parsing, async patterns, error handling
• • Microsoft Foundry portal: projects, deployments, tools
• • Azure resource hierarchy for AI workloads
• • Managed identity vs key-based authentication
• • Exam structure and official study guide walkthrough

Activities

• • Read the official AI-103 study guide. Mark every bullet you cannot explain to a colleague.
• • Verify your Azure subscription is active and create a Foundry project.
• • Deploy an LLM endpoint in Foundry and call it from a Python script that handles the JSON response.
• • Take an ExamOS Mock Exam in Rookie mode. Record your score per domain.

Topics

• • Model selection: LLMs, small language models, multimodal models, code models
• • Foundry Tools vs Foundry Agent Service
• • Deployment options: managed, pay-as-you-go, provisioned
• • CI/CD integration for Foundry projects
• • Quotas, rate limits, scaling, and cost management

Activities

• • Deploy three different model types (LLM, small model, multimodal) in Foundry and document when each is the right choice.
• • Configure diagnostic settings for a Foundry project and confirm logs appear in a Log Analytics workspace.
• • Set up a budget with cost alerts at 80% and 100% in Azure Cost Management for your AI resources.
• • Pass the ExamOS Daily Sprint in Challenger mode twice in a row on the planning domain.

Topics

• • Deploying and consuming LLMs and small models via Foundry SDKs
• • Prompt engineering: system prompts, few-shot, chain-of-thought
• • Model parameters: temperature, top-p, max tokens, stop sequences
• • Evaluating model output for fabrications, relevance, and quality
• • Connecting applications to a Foundry project

Activities

• • Build a Python app that calls an LLM with a system prompt, processes the response, and logs token usage.
• • Implement prompt engineering patterns (few-shot, chain-of-thought) and compare output quality across three different prompts.
• • Write a function that evaluates model output for hallucinations by comparing claims against source text.
• • Pass the ExamOS Daily Sprint in Challenger mode twice in a row.

Topics

• • Document ingestion and indexing: text, images, audio, video
• • Semantic search, hybrid search, and vector search
• • Azure AI Search configuration for grounding
• • RAG ingestion with OCR and layout analysis
• • Retrieval pipeline integration with agent tools

Activities

• • Ingest a set of PDF documents into an Azure AI Search index using vector embeddings.
• • Build a RAG app that retrieves relevant chunks and grounds LLM responses with citations.
• • Compare semantic search vs hybrid search results on the same query set and document the differences.
• • Configure a retrieval pipeline that uses OCR and layout analysis on scanned documents.
• • Pass the ExamOS Daily Sprint in Challenger mode twice in a row.

Topics

• • Agent roles, goals, tool schemas, and conversation memory
• • Function calling and custom tool integration
• • Foundry Agent Service and the Agent Framework
• • Multi-agent orchestration patterns
• • Autonomous and semi-autonomous workflows with approval controls

Activities

• • Build an agent in Foundry with a defined role, tool schema, and conversation memory that persists across turns.
• • Add function calling so the agent can invoke an external API and return structured results.
• • Implement a multi-agent workflow where one agent delegates to another based on task type, with an approval gate for risky actions.
• • Pass the ExamOS Daily Sprint in Challenger mode twice in a row.

Topics

• • Safety filters, content moderation, and guardrails
• • Evaluators for grounding quality, fabrications, and safety
• • Trace logging, provenance metadata, and approval workflows
• • Managed identity, keyless credentials, private networking
• • Agent oversight modes, constraints, and tool-access controls
• • Monitoring: tracing, token analytics, drift detection, latency

Activities

• • Configure content safety filters on a generative app and test with adversarial prompts to verify filtering works.
• • Implement an evaluator that scores grounding quality on a RAG pipeline and log results to Application Insights.
• • Secure a Foundry project with managed identity and private endpoint. Verify access is blocked from outside the VNet.
• • Set up agent oversight with an approval workflow for tool calls that modify external state.
• • Pass the ExamOS Daily Sprint in Challenger mode twice in a row.

Topics

• • Image and video generation from text prompts
• • Inpainting, mask-based edits, and generation controls
• • Multimodal understanding: captioning, visual QA, alt-text
• • Azure Content Understanding pipelines for images and video
• • Language-model text analysis: entities, topics, summaries, structured JSON
• • Sentiment, tone, and sensitive content detection
• • Speech-to-text and text-to-speech for agent interactions
• • Azure Translator and LLM-powered translation flows

Activities

• • Generate images from text prompts and implement a mask-based editing workflow.
• • Build a multimodal app that captions images and answers questions grounded in visual evidence.
• • Extract entities, sentiment, and structured JSON from text using generative prompting and Content Understanding.
• • Add speech-to-text and text-to-speech to an agent so it can handle voice interactions.
• • Pass the ExamOS Daily Sprint in Challenger mode twice in a row.

Topics

• • Full syllabus review across all five domains
• • Time management (120 minutes for 40-60 questions)
• • Case study and lab task technique
• • Responsible AI cross-domain patterns

Activities

• • Take the free official practice assessment again and compare your score to Week 1. Identify remaining weak domains.
• • Take at least 3 ExamOS Mock Exams in Legend mode.
• • Simulate real exam conditions once: no phone, no breaks, 120 minutes, same time of day as your scheduled exam.
• • For every question you miss on a Mock Exam, write one sentence explaining the principle behind the correct answer.
• • Book your exam when your readiness score is stable and above 80, and you have covered every domain at least once.