Twelve common agentic AI interview questions
- What is an AI agent, and how is it different from a chatbot?
- Walk me through the agent loop.
- How does tool or function calling work?
- What is ReAct?
- How do agents plan and decompose tasks?
- What kinds of memory do agents use?
- When would you use multiple agents instead of one?
- How do you evaluate an agent?
- What are the common failure modes, and how do you mitigate them?
- How do you keep an agent safe when it can take real actions?
- How do you control an agent's cost and latency?
- What is the Model Context Protocol (MCP)?
The agent loop
An agent receives a goal, decides on the next step, calls a tool (search, code execution, an API, a database), observes the result, and repeats until it decides the goal is met or a limit is reached. Tool calling works by describing available tools with names, descriptions and parameter schemas; the model emits a structured call, the application executes it and returns the result to the model. ReAct is a prompting pattern that interleaves reasoning and actions, so the model explains why it calls each tool, which improves accuracy and debuggability.
Planning and memory
| Concept | What it means |
|---|---|
| Planning | Decomposing a goal into steps upfront, then re-planning as results arrive |
| Short-term memory | The conversation and tool results in the current context window |
| Long-term memory | Facts or past episodes stored externally and retrieved when relevant |
| Scratchpad | Intermediate notes the agent writes for itself during a task |
| State | Explicit task status the application tracks outside the model |
Single agent or many?
Start with one agent and a good set of tools; it's easier to debug and cheaper. Multiple agents help when sub-tasks need different instructions or tools, when work can run in parallel, or when a separate reviewer agent catching errors improves quality. The costs are coordination overhead, more tokens and harder debugging. A strong answer names a specific reason to split rather than assuming more agents are better.
Failure modes and mitigations
- Loops: the agent repeats the same action. Cap steps, detect repetition, and return partial results.
- Compounding errors: a wrong early step derails everything. Validate tool outputs and re-plan on unexpected results.
- Tool misuse: wrong arguments or the wrong tool. Clear tool descriptions, strict schemas, fewer overlapping tools.
- Prompt injection: a web page or email tells the agent to do something else. Treat tool outputs as untrusted data and limit permissions.
- Runaway cost: budgets per task, smaller models for routine steps, caching.
Evaluating agents
Measure task success rate on a realistic suite of tasks with known outcomes, and look at trajectories, not just final answers: number of steps, tool errors, unnecessary calls, cost and time per task. Include adversarial cases such as injected instructions and tool failures. Re-run the suite on every prompt or model change, because agents regress in surprising ways. Human review of a sample of trajectories catches problems automated checks miss.
Safety and human oversight
Give agents the least privilege they need, run code in sandboxes, require human confirmation for irreversible or external actions (payments, emails, deletions), log every action for audit, and make it easy to stop an agent mid-task. The Model Context Protocol is an open standard for connecting models to tools and data sources through a common interface; it simplifies integration but doesn't remove the need for permission checks on what each tool can do.
Practising
Reading answers is not the same as saying them. Pick ten of the questions above and answer them aloud against a timer; a free AI interview assistant like CoPilot Interview can ask agentic AI questions in a mock interview and flag answers that ran long or lacked an example. Build a small agent with two or three tools and an evaluation set of ten tasks — interviewers ask about real failures you've debugged. See also RAG interview questions.
Disclosure: CoPilot Interview is our product, so weigh what we say about it accordingly. Everything else on this page applies whichever tool you use, or none.
Frequently asked questions
What is agentic AI in an interview answer?
AI systems where a model pursues a goal over multiple steps, deciding which tools to call, observing results and adjusting, rather than producing a single response.
How do you evaluate an AI agent?
Measure task success on a realistic task suite and inspect trajectories — steps, tool errors, cost and time — including adversarial cases, and re-run the suite after every change.
What are common AI agent failure modes?
Loops, compounding errors from a bad early step, wrong tool use, prompt injection through tool outputs, and runaway cost. Step limits, validation, strict schemas, least privilege and budgets mitigate them.
When should you use multiple agents?
When sub-tasks need different instructions or tools, can run in parallel, or benefit from a separate reviewer. Otherwise one agent with good tools is simpler and cheaper.
Practise on real interview questions
Run 145 real interview questions across 10 tracks, answer out loud and see a structured answer to compare with yours. Free, no signup.
Try a practice question