AI Agents Explained: What They Actually Do (And the 5 Worth Trying)

So basically, 2026 is the year AI stopped being a chatbot and started being a coworker. Everywhere you look, somebody’s announcing an “agent” — OpenAI, Anthropic, Google, a dozen startups you’ve never heard of. And if you’re wondering what an agent actually is, you’re not behind. Most of the people selling them would rather you didn’t ask.

Here’s the one-line version: a chatbot answers your question. An agent does your task. You ask a chatbot “what are the cheapest flights to Austin next month” and it tells you. You hand that same job to an agent and it searches, compares, and comes back with the booking page open. One talks. The other works.

Roxie and I have been watching this category all year, and I’ll be honest — most agents are still demos wearing a price tag. But a handful are genuinely useful right now, and the difference between them comes down to one thing nobody advertises: what happens when the agent is wrong. More on that in a minute.

So what is an AI agent, actually?

Think of it like hiring. A chatbot is the really smart friend you text for advice. An agent is the intern you hand a project to. The friend gives you an answer and goes back to their phone. The intern goes off, does the research, builds the spreadsheet, hits a snag, works around it, and comes back with the finished thing.

Under the hood, an agent is just a loop: plan the steps, do a step, check the result, adjust. Chatbots do one step. Agents do the loop. That’s the whole technical difference — everything else is marketing.

And here’s where it gets interesting for normal people: the delegation skill transfers. Everything you learned about giving a chatbot good instructions — being specific, giving it the goal instead of the method, showing it what “done” looks like — matters ten times more with an agent, because an agent will go off and do exactly what you said. Which is great, until it does exactly what you said.

The one thing that makes an agent an agent

Here’s the test I run on every agent that crosses my desk: can it take an action in the real world? Browse a site. Fill a form. Send an email. Move money. If it can only produce text, it’s a chatbot with good branding. If it can click things, it’s an agent.

That power is also the entire risk. An agent with your email login and no supervision is an intern with a company credit card and no manager. The good agents know this — they stop and ask before anything irreversible happens. Sending an email, publishing a doc, spending money: the agent pauses and waits for your yes. The bad ones don’t, and that’s how you wake up to a mess.

So the real question when you’re picking an agent isn’t “how smart is it.” It’s “what does it ask permission for.” Keep that in your pocket — it’s the most useful sentence in this whole article.

1. ChatGPT Agent — the one you might already be paying for

If you’re on ChatGPT Plus at $20 a month, you already own an agent and might not know it. Agent mode arrived in July 2026, and it lives right inside the chat you already use. You describe the task, it browses real websites, fills in forms, reads your files, and produces a finished deliverable — a spreadsheet, a slide deck, a report — instead of a paragraph of advice.

The catch is the meter. On Plus, agent mode shares your plan’s usage allowance, and one long agent session can chew through a real chunk of it — third-party trackers put the Plus cap around 40 agent messages a month, which sounds like a lot until one booking-research task eats six. It’s not on the Free tier or the $8 Go tier at all. So the free-first rule says: great if you already pay for Plus, not a reason to upgrade by itself.

Start with ChatGPT Agent if: you’re already a Plus subscriber and you’ve got multi-step research or form-filling jobs sitting in your backlog.

Roxie’s fine print: the agent is excellent at the doing and mediocre at telling you what it spent. Watch your usage like you’d watch a teenager with a gas card — friendly, but supervised.

2. Manus — hand it a task and walk away

Manus is the most “agent” of the agents. You give it an open-ended task — research a market, build a slide deck, generate a working web app — and it goes off into its own cloud computer and comes back with the finished thing. It’s the contractor, not the assistant. Minimal direction in, finished result out.

The free tier is real: 300 credits a day, no card required. That’s enough to run a proper test, not just a demo. Paid starts at $20 a month for 4,000 credits, and here’s the thing you need to know: everything is priced in credits, and a task’s credit cost is genuinely hard to predict. A deep research run can burn 900 to 1,000 credits — a quarter of the $20 plan — and the price isn’t shown before you hit go. The most common complaint in user reviews isn’t the quality of the work. It’s the bill.

Start with Manus if: you have whole tasks you’d rather delegate than do — research, reports, prototypes — and you can stomach a metered bill.

Roxie’s fine print: run the free tier for a week and watch the credit burn on YOUR tasks before spending a dollar. Billing predictability is the known weak spot — the tool is good, the invoice is a surprise party.

3. Lindy — the teammate that lives where you work

Lindy is the agent for people whose day is email, meetings, and Slack. You message it like a coworker — in a channel, over iMessage, wherever you already are — and it does multi-step work across your actual tools: inbox triage, meeting notes, scheduling, CRM updates. It’s Slack-native in a way the others aren’t, and consequential actions wait for your approval before they go out.

Actually, here’s the Lindy thing you should know: the pricing keeps changing. When I checked in September, the entry plan was $29.99. When I checked this week for this article, the live page showed Free ($50 in credits) and Team at $49.99 a month per user. Same product, new math, six weeks apart. That’s not a red flag about the tool — it’s a yellow flag about planning your budget around it. Check the live pricing page the day you sign up, not the day you read this.

Start with Lindy if: your work lives in email and Slack and you want an agent that meets you there instead of in yet another dashboard.

Roxie’s fine print: Lindy’s own homepage says every active user in your workspace gets billed — including anyone who just @mentions it in Slack. Count your team before you count your savings.

4. Replit Agent — hand it an app idea

Replit Agent is the builder of the bunch. You describe the app or website you want — “a booking page for my detailing business” — and it writes the code, sets up the backend, and deploys it to a live URL. It’s the closest thing to “I had an idea at lunch and shipped it by dinner” that exists right now.

It lives inside Replit, so the free Starter tier gets you in the door and Core at $20 a month is the real starting line. I covered Replit in the coding-tools breakdown, and the agent is the reason it made the cut: for non-developers, this is the lowest-friction path from idea to working software on the internet.

Start with Replit Agent if: you’ve got an app or site idea and “I’d need a developer” has been the excuse for a year.

Roxie’s fine print: the agent writes the code, but you own the result. Read it, test it, and don’t put customer data on anything you haven’t reviewed — that’s not Replit’s job, it’s yours.

5. Genspark Super Agent — the newest one, worth watching

Genspark is the youngest of the five, and it takes a different swing: instead of one model doing everything, it orchestrates multiple models — one researches, another writes, another builds the slides. The output tends to be polished, especially for decks and documents.

It’s credit-based from about $24.99 a month, and honestly it’s the least battle-tested pick on this list. I’m including it because the multi-model approach is where the whole category is heading, and early users are enthusiastic. But “enthusiastic early users” is also what people said about every agent that later died quietly. Try the free tier, keep your expectations honest.

Start with Genspark if: your agent tasks mostly end as documents, decks, or videos, and you like being early.

Roxie’s fine print: newest means least proven. The Relay.app obituary I wrote last month is the cautionary tale — a free tier is a promise, not a contract.

The quick-reference card

Agent	Free tier that matters	Paid from	The catch
ChatGPT Agent	None — needs Plus	$20/mo	~40 agent messages/mo on Plus; sessions eat your allowance
Manus	300 credits/day, no card	$20/mo	Credit costs unpredictable; bill is the #1 complaint
Lindy	$50 in free credits	$49.99/user/mo	Pricing keeps changing; every active user is billed
Replit Agent	Starter free	$20/mo (Core)	You own the code — review before it touches customers
Genspark	Free tier	~$24.99/mo	Newest, least proven of the five

The three rules before you hand anything over

You’ve probably been wondering about this — okay, but how do I actually start without getting burned? Three rules, and they cover almost everything.

First, start with a task you’d be fine redoing. Your first agent run should be something low-stakes — research a purchase, draft a doc, plan a trip. Not your taxes, not a customer email, not anything with money attached. You’re learning what the agent is good at, and tuition should be cheap.

Second, keep the approval gates on. Every agent on this list can be set to ask before it sends, publishes, or spends. Leave those on. An agent that never asks is an agent you haven’t configured yet, not a feature.

Third, check the work like a manager, not a fan. Agents are confident, fast, and wrong in ways that look right. The output is a draft until you’ve verified it — especially numbers, names, and anything you’d sign your name to. The intern metaphor holds: you hired them for speed, you keep the job for judgment.

The verdict: agents are a delegation skill, not a subscription

So what does that actually mean for you? Don’t start by picking an agent — start by picking a task. One annoying, repeatable, multi-step job you hate doing. Then match the agent to it: ChatGPT Agent if you’re already on Plus, Manus’s free tier for the true walk-away test, Lindy’s free credits if your life is email and Slack.

And you know what the real unlock is? The skill transfers. Learning to delegate to an agent — specific goals, clear definitions of done, checking the work — is the same skill as delegating to a person. The agents will get better every quarter. The people who know how to hand them work are the ones who’ll get the value.

That’s the whole game with agents. Not the smartest model. The clearest instructions.