ung-dung Intermediate

What is an AI Agent?

An AI system that can plan, use tools, and take multi-step actions to complete a goal — not just answer a single question.

Updated: May 2, 2026 · 2 min read

An AI agent uses an LLM as its “brain” to make decisions, call external tools, and take multiple actions in sequence to complete a complex goal.

How agents differ from chatbots

Chatbot	Agent
Answers one question at a time	Loops through multiple steps
Generates text only	Can take actions (call APIs, write files, drive a browser)
Needs the user to direct each step	Plans on its own

Real-world agents (2026)

Coding agents

Claude Code, Cursor, Devin: tell them “fix bug X” and they read code, edit files, run tests, commit.

Research agents

Perplexity Pro, OpenAI Deep Research: ask for “an EV market report on ASEAN” and they search the web, read 50 sources, synthesize.

Browser agents

Computer Use (Anthropic), OpenAI Operator: control mouse and keyboard to book tickets, fill forms, operate apps.

Personal assistants

Schedule meetings, read email, plan trips — integrating with Calendar, Gmail, flights.

A simple agent loop

┌──────────────────────────────────┐
│  LLM (brain) — Claude/GPT        │
│                                  │
│  Loop:                           │
│    1. Read task + state          │
│    2. Think (Chain of Thought)   │
│    3. Pick a tool + parameters   │
│    4. Call tool, get result      │
│    5. Go to step 1 if not done   │
└──────────────────────────────────┘
         ↓ tools
   [search] [file_read] [code_run] [http_call] ...

Why agents are exploding (2025-26)

Three things converged:

Tool use became reliable (better function calling)
Reasoning models (Claude 4.7, GPT-5) think much better
MCP standardized how agents connect to tools

→ Agents are no longer demos — they’re production-ready for many use cases.

Current limitations

Compounding errors: 95% per step × 10 steps = 60% end-to-end success
Cost: agents call the LLM many times → 5-50× more expensive than chatbots
Latency: many round trips → slow
Hard to debug when they go wrong

When to use an agent

✅ Repetitive tasks with a clear workflow ✅ Time-consuming work that doesn’t need genius reasoning ✅ Users tolerate tens-of-seconds to multi-minute latency

❌ Need real-time response ❌ Need 99.9% accuracy (medicine, finance) ❌ Tight budget