What Is an AI Agent? A Plain-English Guide

An AI agent is software that does work for you, step by step, instead of just answering questions. Where a chatbot waits for your next prompt, an agent takes a goal — "compare these three options and book the cheapest" — plans the steps, uses tools like web search and code execution, checks its own work, and reports back. This guide explains what agents actually do in late 2026, what they cost, where they fail, and how to start safely.

Coding agents that read, change and test whole codebases lead the field. Photos: Pexels (free license).

Quick answer: coding agents are the most mature (and worth trying today); research agents are excellent with verification; office agents save real time on drafts and summaries; browser agents that click around the web for you are the least reliable — supervise them. Start with read-only tasks, tight permissions, and review everything.

Contents

• Chatbot vs agent: the difference
• Coding agents · Research agents · Office agents · Browser agents
• Why agents work now · Risks · Safe-start ladder · Cost
• Picking your first agent · FAQ

Chatbot vs agent: the difference

A chatbot answers: you type a question, it types an answer, context resets every turn. An agent acts across many steps without you: it decomposes your goal, calls tools (search, file access, code execution, calendars, APIs), observes results, adjusts the plan, and only returns when done — or stuck. Concretely: a chatbot explains how to compare flights; an agent searches, filters, compares and drafts the booking while you drink coffee. The shift from 2024 to 2026 is exactly this: assistants stopped only talking about tasks and started finishing them.

Coding agents — the most mature category

Give a coding agent repository access and a task ("add password reset with tests") and it reads the codebase, writes the feature, runs the test suite, fixes failures and opens the change for review. Leaders include Claude Code (strong at review and multi-file refactoring), OpenAI's Codex agent (included even in the $20 ChatGPT Plus plan), Cursor (~$20/month) and GitHub Copilot's agent mode. Professional developers now routinely start tasks with an agent and review the diff — the human job moved up a level from typing code to specifying and verifying it.

Verdict: if you write software and have not tried an agent on a real branch yet, this is the highest-value hour you can spend this week. For choosing the underlying assistant, see our Claude vs ChatGPT comparison.

Research agents — excellent, with verification

A research agent turns "should we expand to Portugal?" into a planned investigation: it searches dozens of sources, follows citations, compiles numbers into tables and returns a cited report instead of ten blue links. ChatGPT's Deep Research (about 10 runs a month on Plus, far more on Pro tiers), Perplexity's research modes and similar offerings from Gemini and Claude all work this way. Quality varies with question specificity — vague goals produce vague reports. Verdict: demand sources in the output, then spot-check the three claims that matter most. Our free alternatives guide covers the $0 research options.

Office and workflow agents — quiet time-savers

The least glamorous, most-used category: agents living inside Word, Gmail, Notion and similar tools that draft emails, summarize 200-message threads, prepare meeting briefs and keep docs current. Microsoft's Copilot agents, Google's Workspace agents and Notion's AI agents (full access on its Business plan) all target this. Savings are real but boring — roughly an hour a day for heavy email/calendar users. Verdict: adopt where you already work; do not buy a new suite just for the agent.

The leap: from answering questions to finishing tasks. Photos: Pexels (free license).

Browser and computer agents — promising, supervise closely

These operate websites and apps like a human would: clicking, filling forms, comparing checkout flows. OpenAI's Operator-style agent mode (higher tiers) and similar browser agents can complete real multi-site errands — and also misread pages, loop on login screens and confidently do the wrong thing quickly. Reliability is visibly lower than the categories above. Verdict: useful for low-stakes errands with you watching; never hand one your credit card or inbox unsupervised.

Why agents work now (and flopped before)

Three things changed: longer context (current models hold up to ~1M tokens, so an agent remembers a whole project), reliable tool use (models learned to call search, code and APIs instead of just describing them), and multi-step reasoning (explicit thinking modes that plan before acting). Scripted "bots" of the 2010s broke on step three of anything; 2026 agents replan when step three fails. They still fail — just less often, and more recoverably.

The risks, stated plainly

• Compounding errors: a wrong assumption in step 2 poisons steps 3–20. Review agent output like a new hire's first week — trust is earned per task.
• Permissions are the blast radius: an agent with inbox, file and payment access can make expensive mistakes fast. Grant the minimum scope that completes the task.
• Prompt injection: agents read web pages and files, and malicious text in those sources can hijack their instructions. Do not point agents at untrusted content with write permissions attached.
• Cost and limits: agentic runs burn far more compute than single prompts — watch usage caps on paid plans and per-token API bills.
• Confident wrongness: agents present finished-looking work. Finished-looking is not finished-correct. Verify before publishing, paying or sending.

The safe-start ladder

1. Read-only research: summaries, comparisons, code explanations. Zero permissions, full review.
2. Drafts you send: emails, posts, proposals — agent drafts, you press send.
3. Code on a branch: agent commits to a branch, tests run automatically, you merge.
4. Reversible actions: filing, scheduling, organizing — undoable in one click.
5. Only then: broader access, one capability at a time, with spending/permission caps where offered.

Rule of thumb: never grant an agent a permission whose worst case you cannot afford.

What agents cost

Most bundled agents ride inside existing subscriptions: Codex in ChatGPT Plus ($20/month), Copilot agents in Microsoft 365 tiers, Notion AI on its Business plan (~$20/seat/month), Claude Code within Claude plans (Pro $20/month, heavier Max tiers above). Standalone power tiers (like ChatGPT's $100–$200 Pro plans) buy dramatically higher agent usage for professionals who run them all day. API-built agents bill per token — cheap to prototype, worth metering in production. Start on what you already pay for; upgrade when caps block daily work.

Picking your first agent (by job)

• You code → Claude Code or Codex on a real branch this week.
• You research → a Deep Research run on a question you already answered — compare quality against your own work.
• You drown in email → your suite's built-in agent (Copilot, Gemini, Notion AI) for one week of summaries and drafts.
• You organize knowledge → an agent that files and summarizes into your notes system — and read our Notion vs Obsidian guide if AI-written work needs a home.
• You are just curious → any free chatbot's agent mode on a read-only errand. Our free alternatives guide lists the $0 doors.

How we wrote this

From product documentation, official pricing pages and established professional reviews of the named tools, current to late September 2026 — not vendor launch demos, which always show the happy path. Agent capabilities move monthly: model names, limits and tier contents will drift. We describe failure modes bluntly because they are the load-bearing facts of this category.

Frequently asked questions

What is the difference between an AI agent and a chatbot?
A chatbot answers one prompt at a time; an agent pursues a goal across many steps, using tools (search, code, files, apps) and adjusting its plan along the way.

What is the best AI agent for beginners?
A research agent on a read-only task — no permissions to grant, nothing to break, and you learn to judge output quality before trusting agents with actions.

Are AI agents safe?
As safe as their permissions. Read-only agents are nearly risk-free; agents with inbox, file or payment access need tight scopes and human review. Never combine untrusted sources with write permissions.

Do AI agents cost extra?
Usually they are bundled: Codex with Plus, Copilot agents with Microsoft 365, Notion AI with Business. Heavy all-day agent use is what the $100–$200 power tiers exist for.

Will agents replace virtual assistants or junior developers?
They already absorb specific junior-level tasks (first-draft research, boilerplate code, inbox triage) — with a human reviewing. The job shifts to specifying goals precisely and verifying results, which is itself a skill worth building now.

Keep exploring: Claude vs ChatGPT · 10 free ChatGPT alternatives · Notion vs Obsidian · About this site

Agent features change monthly — confirm tier contents on each provider's official site.