Agentic AI earns its complexity on tasks that genuinely need multi-step, dynamic decision-making — not on tasks a single prompt or a fixed workflow already handles well.
Research & Information Gathering
An agent tasked with "research competitor pricing for product X" can decide which sources to check, read results, follow up on gaps, and synthesize a report — a task where the right steps depend entirely on what's found along the way. See Research Agent.
Data Analysis
An agent with database access can be asked an open-ended question ("why did revenue drop in the North region last month?"), decide which queries to run, interpret results, and run follow-up queries based on what it finds — rather than a fixed, pre-written report. See SQL Agent.
Customer Support (Beyond FAQs)
An agent that can look up a specific order, check a return eligibility policy against that order's details, and — for anything beyond a routine case — escalate to a human, rather than a fixed FAQ chatbot. See Customer Support Agent.
Coding Assistance
An agent that can read a codebase, run tests, interpret failures, and iterate on a fix — going beyond single-shot code generation into a genuine debug loop. See Coding Agent.
Document Processing
An agent that decides how to handle a document based on its type — routing a scanned form to OCR, a structured PDF to direct parsing, and flagging unrecognized formats for human review. See Document Agent.
What These Have in Common
Every example above shares a trait: the correct sequence of steps genuinely can't be fully known in advance — it depends on what's discovered along the way. That's the real signal for reaching for an agent instead of a simpler pattern.
Common Mistakes
- Applying agentic patterns to tasks with a clearly fixed, known sequence of steps — added cost and unpredictability for no real benefit
- Skipping evaluation before shipping an agent to production because "it worked in the demo" — agent behavior on edge cases needs deliberate testing (see Agent Evaluation)
Interview Relevance
"Give a real use case where an agent is clearly the right architecture, and one where it's overkill" — tests practical judgment, not just familiarity with the concept.
Practice Question
Pick one use case above and list two tools the agent would need access to, and one action that should require human approval before executing.
