Agentic AI describes systems where an LLM doesn't just respond once — it runs in a loop, deciding which actions or tools to use, observing the results, and deciding what to do next, until it reaches a goal or a stopping condition.
The Core Shift: From One Response to a Loop
Plain LLM call:
prompt → model → one response → done
Agentic loop:
goal → model decides action → action runs → observe result
→ model decides next action → ... → final responseThe difference isn't the model — it's the same underlying LLM either way. The difference is the system built around it: does it call the model once, or does it let the model make a sequence of decisions based on what happened after each step?
A Concrete Example
Ask "What's the weather in Delhi and should I carry an umbrella?" to two different systems:
| Plain LLM | Agent with a weather tool |
|---|---|
| Generates a plausible-sounding answer from training data — which may be outdated or simply guessed, since it has no live weather access | Decides it needs current data → calls a weather API tool → reads the actual result → generates an answer grounded in that real data |
See How AI Agents Work for the full mechanics of that loop.
What "Agentic" Does Not Mean
It does not mean the system is fully autonomous, unsupervised, or free of constraints. In practice, well-built agentic systems operate within explicitly defined boundaries: a fixed set of available tools, permission checks, iteration limits, and — for high-impact actions — human approval steps (see Human-in-the-Loop). "Agentic" describes a decision-making pattern, not a claim about unlimited independence.
Why This Matters for Engineering Decisions
Not every LLM feature needs to be agentic. A single-turn Q&A feature, a fixed data-extraction pipeline, or a deterministic report generator often works better as a plain LLM call or a fixed workflow — simpler to test, cheaper to run, and easier to make reliable. Agentic patterns earn their added complexity specifically when the task genuinely requires dynamic, multi-step decision-making that can't be reasonably predetermined.
Common Mistakes
- Calling any chatbot an "agent" — a single-turn conversational LLM with no tools and no decision loop is not agentic; see Agent vs Chatbot
- Overclaiming autonomy — describing agentic systems as fully independent decision-makers with no oversight, which misrepresents both the technology and its risk profile
- Reaching for an agent when a deterministic workflow would be simpler, cheaper, and more reliable for the actual task
Interview Relevance
Q: "What makes a system 'agentic' rather than just an LLM-powered feature?" A strong answer centers on the decision loop — the system observes results and decides its next action dynamically, rather than following a single fixed prompt-response or a predetermined sequence of steps.
Practice Question
A customer support bot answers FAQs from a fixed document using a single retrieval-then-generate step. Is this agentic? Justify your answer using the definition above.
