What is an AI Agent?

LLM + tools + loop: the model plans, acts, observes, and repeats until the task is done.

Interview tip Contrast: chatbot = one-shot. Agent = multi-step with tools (search, code run, API calls). Mention human-in-the-loop for risky actions.

① What you must know (30 sec)

An AI agent is an LLM that runs in a loop: it receives a goal, decides the next action (often a tool call), executes it, observes the result, and repeats until it produces a final answer or hits a limit. Unlike a one-shot chatbot, agents can search the web, query databases, run code, and call APIs — with variable cost and risk.

Analogy: A chatbot is a consultant who answers from memory; an agent is an intern with a laptop who can look things up, send emails, and run spreadsheets until the task is done.

② How it works

Agent loop

User goal→LLM plans action→Tool executes→Observation→Final answer or repeat

User goal — natural language task with success criteria
LLM planner — chooses tool + arguments or returns final answer
Tool runtime — validates args, executes allowlisted function
Observation — tool output appended to conversation history
Termination — final answer, max_steps, timeout, or human approval

ReAct (Reason + Act) is the most common pattern: the model explicitly reasons before each tool call.

③ Step-by-step (hands-on)

Step 1 — Define the goal and success criteria

"Summarize open Jira tickets for team X and post to Slack" — clear done state prevents infinite loops.

Step 2 — Inventory tools

List capabilities the LLM cannot do alone: search, SQL, calendar, send_message. One tool = one well-documented function with JSON schema.

Step 3 — Write planner prompt

Explain available tools, output format (JSON action or final answer), and rules: "Never guess account IDs; use lookup tool."

Step 4 — Implement the loop

Call LLM → parse tool call → execute → append result → repeat. Cap at 5–15 steps and 2–5 minute timeout.

Step 5 — Add guardrails

Allowlist tools, validate parameters, require approval for writes, sandbox code execution, audit log every action.

Step 6 — Observe and evaluate

Trace each step (LangSmith, OpenTelemetry). Measure task success rate and cost per run on a test suite.

④ Code / config patterns

Pattern	Description	Best for
ReAct	Reason then act with tools	General-purpose agents
Plan-and-execute	Plan all steps upfront, then run	Predictable multi-step workflows
Router	Classify intent → specialized sub-agent	Multi-domain assistants
Human-in-the-loop	Pause for approval on risky actions	Finance, healthcare, prod changes

while not done and steps < MAX_STEPS:
    response = llm(messages, tools=tool_definitions)
    if response.finish_reason == "final":
        return response.text
    args = validate(response.tool_call)
    result = TOOL_REGISTRY[response.tool_name](**args)
    messages.append({"role": "tool", "content": result})
    steps += 1
raise TimeoutError("Agent exceeded step limit")

⑤ Production & pitfalls

Pitfall	Why it hurts	Fix
Unbounded loops	Runaway cost and latency	max_steps, timeout, token budget per run
Arbitrary code execution	RCE, data exfiltration	Sandbox, allowlist commands, no shell by default
Tool description ambiguity	Wrong tool chosen repeatedly	Clear names, examples, negative cases in schema
No argument validation	SQL injection, bad API calls	Schema validation before execution
Agent for simple FAQ	10× cost vs one RAG call	Use single LLM call when no tools needed
Opaque failures	Cannot debug bad runs	Structured traces per step with inputs/outputs

Production tips:

Separate read-only and write tools; gate writes behind approval
Idempotent tools where possible so retries are safe
Per-tenant tool permissions and data isolation
Kill switch to disable agent features without full outage

⑥ Interview / on-the-job Q&A

Question	Answer
Agent vs chatbot?	Chatbot: one LLM call, no side effects. Agent: multi-step loop with tool execution.
What is a tool?	A function the LLM can invoke — search, SQL, API — described by name, description, and JSON schema.
What is max_steps?	Hard cap on loop iterations to prevent runaway cost and infinite retries.
When not to use agents?	Simple Q&A, single retrieval, or deterministic workflows better handled by code.
What is human-in-the-loop?	Pause before irreversible actions (send email, charge card) for human approval.
How debug agents?	Step-by-step traces: prompt, tool chosen, args, result, and final answer per run.

⑦ Tools & ecosystem

Frameworks: LangGraph, CrewAI, OpenAI Assistants, Anthropic tool use
Tracing: LangSmith, Arize, Weights & Biases
Sandbox: E2B, Docker, Modal for code execution
Protocols: MCP for standardized tool connections

agentstoolsreactguardrailstracing

⑧ Revision checklist

Clear success criteria defined before building loop
Tools have JSON schemas with descriptions and examples
max_steps and timeout configured
Write/destructive tools require approval
All tool calls audit-logged with user and timestamp
Argument validation before execution
Traces captured for every run
Eval suite of 10+ multi-step tasks
Cost per successful task measured and acceptable