What Is Agentic AI?

Chatbots answer questions. Agents do things. That one distinction changes what's possible with AI — and what risks come with it.

Diagram showing an AI agent loop: receive task, plan steps, take action, check result, repeat — versus a chatbot that only responds
Agentic AI vs chatbots — the core difference is action versus response

The Core Difference

A regular AI chatbot responds. You send a message, it replies, you do the work with what it gave you. Every interaction requires you.

An AI agent acts. You give it a goal. It figures out the steps, takes them, checks whether they worked, adjusts, and continues — until the task is done or it gets stuck.

Same underlying AI technology. Completely different relationship to the work.

What "Acting" Actually Means

An AI agent can use tools. The tools determine what it can act on. Common tools agents use:

  • Web browsing — visit URLs, read pages, follow links
  • Code execution — write Python or JavaScript and actually run it, see the output, fix errors
  • File operations — create, read, edit, and delete files
  • API calls — interact with external services (check weather, post to Slack, send email)
  • Computer control — control a mouse and keyboard on a real computer screen

Not every agent has all of these tools. What an agent can do depends entirely on which tools it's been given access to.

Real Examples of What Agents Can Do Today

Works well now

Deep research reports

Give an agent a topic. It browses 15–20 sources, pulls the relevant information, structures it, cites sources. Produces a report that would take a human 3–4 hours. Agent does it in 20 minutes.

Works well now

Data analysis pipelines

Upload a spreadsheet. Ask for analysis. The agent writes the analysis code, runs it, checks for errors, fixes them, and delivers charts and a summary — without you writing a line of code.

Works, needs review

Web form filling and data extraction

An agent can browse websites and extract structured information — prices, contacts, listings. Results need verification. It makes mistakes on edge cases.

Works, needs review

Code generation across a project

Tools like Cursor AI can understand a full codebase and make changes across multiple files. Works well for defined tasks; requires review before committing anything.

Not reliable yet

Autonomous financial transactions

Agents should not handle financial transactions unsupervised. Too much can go wrong, and mistakes are costly and sometimes irreversible.

Not reliable yet

Long unsupervised workflows

Anything requiring 20+ steps and judgment calls along the way. Agents tend to drift, make wrong assumptions, and compound errors. Check in frequently.

Agentic AI Tools Available in Canada

  • Manus AI — general-purpose agent, web browsing + coding + file management. Free tier + paid (~$55 CAD/month).
  • OpenAI Operator — browses websites to complete tasks. Included with ChatGPT Pro (~$280 CAD/month). Still early.
  • Claude computer use — Claude can control a computer screen. Available via API for developers.
  • Cursor AI — agentic coding. Understands your codebase, makes multi-file changes.
  • Microsoft Copilot — integrates into Word, Excel, Outlook. Can take actions across your Microsoft 365 documents. Requires Microsoft 365 Business plan.

The Safety Question

Agents that can take actions raise a question regular chatbots don't: what happens when the agent does something wrong?

With a chatbot: it tells you incorrect information. You verify, discard it, move on.

With an agent: it might delete a file, send an email, submit a form, or make a purchase. Some actions cannot be undone.

The practical rule: never give an AI agent access to anything you would not be comfortable with it destroying or spending. No access to your banking. No access to your email without strict limits. No access to production systems unsupervised. Start with read-only permissions and expand carefully.

Where This Is Going

Every major AI lab is investing heavily in agents. The next 12–24 months will see agents that can reliably handle more complex, longer-running tasks with less supervision.

The most likely near-term use cases for most people: research tasks, scheduling, document drafting and editing, and software development assistance. All of these are already partially working. They'll become more reliable as the underlying models improve and as agent frameworks get better at error-checking and recovery.

Bottom Line

Agentic AI is the difference between a knowledgeable assistant who tells you what to do and one who does it for you. It's more powerful and more risky. The tools that exist today work well for research and coding tasks. Everything else needs human oversight at each step.

Try Manus AI on the free tier with a research task. That's the fastest way to understand what agents can actually do.