Back to Q&A board 
The most common agentic patterns discussed in interviews are:
Router – Classifies requests and selects the appropriate model, tool, agent or workflow.
ReAct / Tool-use loop – The model selects a tool, observes the result and decides the next action.
Planner–Executor – A planner decomposes the objective; an executor performs and verifies each step.
Sequential workflow – Specialized stages run in a controlled order, such as retrieve → analyze → generate → validate.
Orchestrator–Workers – A coordinator delegates independent tasks to specialized or parallel workers and aggregates their results.
Reflection / Critic – A reviewer evaluates an output and requests bounded improvements.
Agentic RAG – The system determines whether retrieval is necessary, rewrites queries, searches iteratively and verifies supporting evidence.
Memory-augmented agent – Uses short-term state and selectively retrieved long-term memory.
Human-in-the-loop – Requires approval for sensitive or irreversible actions such as payments, deployments or external communication.
Event-driven agent – Suspends and resumes long-running work when timers, messages or external events arrive.
Multi-agent collaboration or debate – Multiple agents propose, challenge or validate solutions before synthesis.
In a strong interview answer, do not present autonomy as the objective. Explain that production systems normally use a deterministic workflow engine, with LLM reasoning restricted to controlled steps. Add typed tools, permissions, iteration and cost limits, idempotency, retries, observability, evaluation and human approval.
For a deeper architectural treatment, I recommend my white paper on SSRN: read the paper. It provides useful preparation for discussing how agentic patterns should be structured and controlled in production systems.