AI-500 practice test

Decomposing goals into workflows, agents and tools, designing control loops and human-in-the-loop steps, specifying agent personas, scopes and autonomy levels, choosing communication protocols, designing short-term and long-term memory including shared context, matching tasks to model families, and specifying the integration, state, compute, observability and monitoring components a production multi-agent system needs.

Advanced prompt engineering and dynamic context injection, context accumulation, retrieval and compaction, memory strategies across agents, multi-agent retrieval architectures, tool ecosystems including function calling and MCP servers built on Functions, Logic Apps and API Management, tool error handling and result validation, and the orchestration itself: hub-and-spoke, sequential, parallel, peer-to-peer and orchestrator-subagent, with caching, concurrency control and middleware.

Human review processes and evaluations for memory, knowledge, tools and prompts, optimising task duration against rate limits, diagnosing context window failures such as sliding-window amnesia and summary drift, continuous improvement through model-as-a-judge frameworks and feedback loops, monitoring agent health, workflow failures, trace correlation and drift, token and cost management, and tracing in Foundry.

Identity-based resource access and per-agent scoping, authentication flows including on-behalf-of and OAuth, secrets and key rotation in Key Vault, red teaming before release, guardrails intervening on inputs, tool calls, tool responses and outputs, custom guardrails and their testing, and the release work: environments, blue/accent and canary rollout, rollback, test strategy and continuous delivery.