How to use AI agents to get real work done, not just chat

Executive overview

Most people use AI as a chat tool — ask a question, copy the answer, go do the work yourself. That approach is already obsolete. Agentic AI operates at a higher level: it plans, reasons, and executes entire tasks autonomously, including opening browsers, writing code, and managing files.

The shift required is from doer to director. Your job is to define the outcome and review the output — not to execute.

The director mindset unlocks the real value of AI: not faster answers, but completed work.

The three levels of AI

• Chat — conversational tools like ChatGPT, Claude, Gemini; you prompt, it replies
• Automation — AI-powered workflows via tools like Make, Zapier, or N8N
• Agents — AI that thinks, plans, and executes full tasks without hand-holding; you don't need level two to reach level three

How to direct AI effectively

• Define a clear outcome before starting — what is the specific result you want?
• Provide a format, template, or example output so the AI knows what "done" looks like
• Give feedback and ask it to save that feedback; it builds a working style over time
• Never tell it how to do the work — start with the end, let it figure out the path
• Treat it like an intern: review, correct, iterate

Choosing the right tool

• Manus AI — best for business owners needing research, content, and general task completion
• Claude Cowork (creatives) — writing, design, ideas; manages files, opens tabs, runs full projects on your computer
• Claude Cowork (developers) — bug fixes, test files, parallel codebase work; the tool used internally by AI company developers
• OpenClaw — fully automated personal assistant with memory; more technical setup, higher risk (one example: it autonomously purchased a $3,000 course)
• Pick one tool and go deep; switching between many builds mastery of none

Manus in action: a real example

• Prompt: research the top three digital agencies in a niche in Canada — pricing, features, strengths — then build a one-page website summarising findings
• Manus builds a task list, does the research, writes the code, and produces the site without any keyboard input
• Iteration: ask it to add a testimonials section — it updates the site in minutes
• Integration: it posted the result to Slack as the user, emailed a link to a colleague, and was set to monitor the Slack thread for feedback
• Parallel runs: multiple agent tasks can run simultaneously while you do other work
• Mobile: the full workflow is accessible and operable from a phone

Pro tips for building the habit

• Stay in the tool — resist copying output and doing the next step yourself; let the agent do it
• The more you delegate within the tool, the better both you and the AI get at working together
• Start with one task you do every week that takes significant time; have the agent do it today