Q&A

What are AI Agents?

By AlgoInsight Team2026-06-18algorithmsdatabases

Overview

The fundamental insight behind agents is that an LLM can not only generate text but also decide what to do next - reason about a goal, choose a tool, interpret the result, and loop. This ReAct (Reason + Act) loop lets a single LLM orchestrate a sequence of actions that would otherwise require explicit programming.

graph TD
    A[User goal] --> B[Agent: LLM-based reasoner]
    B --> C{Which tool to use?}
    C --> D["Web search"]
    C --> E["Code execution"]
    C --> F["Database query"]
    C --> G["API call"]
    D & E & F & G --> H[Observation / result]
    H --> B
    B --> I{Goal achieved?}
    I -->|yes| J[Return final answer]
    I -->|no| C
    style B fill:#1e1b4b,color:#fff

The ReAct Loop

The core agent pattern interleaves Reasoning and Acting:

Thought: The user wants the weather in London. I should search for it.
Action: web_search("London weather today")
Observation: "London: 18°C, partly cloudy"
Thought: I have the answer.
Final Answer: It's 18°C and partly cloudy in London today.

At each step the LLM decides whether to call a tool or conclude. This loop repeats until the task is done or a limit is reached.

Tool Use

Tools are functions the agent can call. They're described to the LLM via a schema:

{
  "name": "web_search",
  "description": "Search the internet for current information",
  "parameters": {
    "query": {"type": "string", "description": "The search query"}
  }
}

The LLM selects the tool and arguments; your code executes it and returns the result. Modern APIs (OpenAI function calling, Anthropic tool use) handle the serialization.

Memory Types

Multi-Agent Systems

Complex tasks can be split across multiple specialized agents:

• Orchestrator - decomposes the goal and delegates.
• Subagents - specialists (researcher, coder, critic, executor).
• Tool agents - single-purpose wrappers around a tool.

Frameworks like LangGraph, AutoGen, and CrewAI formalize multi-agent workflows with defined roles, message passing, and state management.

Popular Agent Frameworks

Key Challenges

• Reliability - agents can take wrong actions or get stuck in loops; human-in-the-loop checkpoints help.
• Context window - long task histories fill the context; need summarization or external memory.
• Error recovery - the agent must detect and recover from tool failures.
• Safety - agents with write access (delete files, send emails) can cause irreversible harm; least-privilege and confirmation steps are critical.
• Cost - multi-step loops with large models are expensive; optimize with smaller models for subagents.

Why This Matters in Interviews

AI agents are the dominant product paradigm in 2025-26. Interviewers test: the ReAct reasoning loop, tool use / function calling, the four memory types, multi-agent patterns (orchestrator + subagents), and practical challenges (reliability, safety, cost). Distinguishing a simple RAG pipeline (retrieve then generate, one shot) from an agentic loop (iterative, tool-using, stateful) is the key conceptual line.

Common Pitfalls

• Designing agents without human-in-the-loop for irreversible actions (deletes, sends, purchases).
• Unlimited retry loops without a stopping condition - infinite loops waste money and time.
• Putting too much responsibility in one mega-agent instead of decomposing into specialized subagents.
• No observability - logging every thought/action/observation is essential for debugging.
• Ignoring context window limits in long-running tasks.

References

• Yao et al., "ReAct: Synergizing Reasoning and Acting in LLMs" (2022): https://arxiv.org/abs/2210.03629
• LangGraph docs: https://langchain-ai.github.io/langgraph/

More answers