🤖 AI agents for beginners: what they actually are (and aren't)

The plain-English definition

An AI agent is software you give a goal to — and instead of just answering you, it works out the steps itself and then actually does them, using the apps, files and websites you already use, checking back with you before anything risky.

That last part matters. A chatbot tells you how to book a flight. An agent could actually search the flights, compare the prices, fill in the booking form and ask you to confirm before it pays.

Agent vs chatbot — the real difference

• A chatbot RESPONDS. An agent ACTS.
• Chatbot: you do the work after it answers. Agent: it does the work, then tells you it's done.
• Chatbot: no access to your files or apps. Agent: connected to your email, calendar, files, browser.
• Chatbot: one answer at a time. Agent: a chain of steps toward a goal.
• Honest caveat: the line blurs. A weak agent is basically a chatbot with tools bolted on — and plenty of 'agents' marketed today are just that.

Real tools you can actually try

• ChatGPT — has an agent mode ('ChatGPT Work'). Free tier includes it on desktop. Plus is about R399/month in South Africa.
• Claude — free tier plus Pro (around R300–350/month) with its 'Cowork' agent features.
• Google Gemini — 'Deep Research' is arguably the best FREE first experience with an agent. Google AI Plus is around R105/month in SA.
• Perplexity — its Comet agent browser has been free since March 2026. Pro is about $20/month.
• Zapier Agents — free tier (400 tasks/month), no card needed. Great for automating between your existing apps.

What agents can actually do today

This is where honesty matters, because there's a lot of hype. As of 2026, agents are genuinely good at research, drafting, summarising your inbox, and coding (with a human reviewing). They are NOT yet reliable at long, unattended, multi-step work.

Gartner reported in April 2026 that only about 17% of organisations had actually deployed agents, and full autonomy is still 'not ready'. That's the honest state of play: the demos are impressive, the dependable reality is narrower.

The misconceptions to drop

• "It's just a smarter chatbot" — no, the difference is that it acts, not just answers.
• "It's an autonomous employee" — no, it needs supervision and checkpoints.
• "It's free" — agents burn credits fast; costs meter per step.
• "More autonomy is better" — no, more autonomy means more ways for it to go wrong unattended.
• "If it says it did X, it did X" — verify. Agents can report success without succeeding.

The real risks (so you don't get burned)

• Cost — agents can loop and burn through credits. Set limits.
• Mistakes — there are documented cases of agents deleting real data (including a production database) because they were given too much access and no confirmation step.
• Security — the OWASP 'Agentic Top 10' exists for a reason: prompt-injection attacks can trick an agent into doing things you didn't intend, including making payments.
• Privacy — don't give an agent access to sensitive accounts until you trust it.

How to take a safe first step

Start on a free tier. Give it ONE small, reversible task — summarise these emails, research this question, draft this document. Do not connect it to your bank, your main email, or anything you can't undo. Set permissions to 'always ask'. And verify everything it claims it did.

Scale up only after it's proven reliable on the small stuff. That's the whole discipline: small, reversible, supervised — then more only when it earns it.