Agentic AI Concepts and Architecture
What this domain tests
Agentic AI Concepts and Architecture tests whether you can separate an agent from a chatbot and describe the components that make autonomous work safe: model, instructions, tools, memory, policy, and orchestration. Expect questions on planning loops, single- versus multi-agent designs, and where human approval belongs.
Core rules
- An agent is a goal-directed system that observes, decides, acts through tools, and checks results. A generative model alone is not an agent.
- The common control loop is reason-act-observe: plan a step, call a tool, read the result, update state, and repeat until a stop condition or escalation.
- Memory is not one thing. Short-term working memory holds the current task; long-term memory persists preferences or case history and needs retention rules.
- Orchestration can be a single agent with many tools, a supervisor with specialist agents, or a deterministic workflow that calls models at fixed points. Choose the least autonomy that meets the risk.
- Every loop needs bounds: maximum steps, timeouts, budget, and a defined escalation path.
Worked example
An invoice-dispute agent receives a customer email. It extracts invoice number and reason, retrieves contract terms, calls the ERP API for invoice status, and drafts a response. If the disputed amount exceeds the approval limit, it must stop and route to a finance specialist rather than issue a credit. The reasoning is architectural: retrieval supplies policy, the ERP tool supplies authoritative state, and the threshold is a policy guardrail, not a model judgment. The plan is bounded because the credit action is irreversible.
Traps
Do not describe RAG as agency; retrieval answers a question but does not decide or act. Do not equate more autonomy with a better design. Do not let the model choose its own approval threshold. Do not omit idempotency for actions that can be retried.