Lesson 17. Agentic AI: multi-step tasks
Agentic AI is when the model doesn't just answer a question but autonomously performs multiple steps in sequence. It creates a plan, uses tools, checks results, and continues until reaching the goal. This is fundamentally different from regular chat — and in this lesson you'll learn how to use this capability.
Topic breakdown
An AI agent doesn't just respond — it autonomously plans and executes a chain of actions. This can include search, computation, file reading, API requests, and other operations.
ChatGPT Operator, Claude computer use, Cursor, Replit Agent — these tools already exist. For regular business, Make.com, Zapier, or n8n allow connecting AI to automatic workflows.
Strength and weakness in one place: if the agent makes a mistake, the chain continues, and fixing things at the end is difficult. That's why checkpoints at each step and the instruction 'if unsure — stop' are critically important.
Agentic workflows are most effective for repeating, clearly structured tasks that can be performed without constant human involvement: data collection, report preparation, sending messages.
What you'll learn
- understand the difference between agentic and regular AI
- write instructions for agents with multi-step tasks
- envision a simple AI workflow in Make.com or Zapier
- identify control points and agent error handling
Lesson plan
Agentic AI vs regular chat: what's the difference?
Regular chat: one question — one answer. AI agent autonomously plans and executes multiple steps, uses tools, and strives toward a final goal.
How to write instructions for an agent?
Goal, list of tools, sequence of steps, and stop points — four elements form the basis of an agentic prompt.
Automation platforms
Make.com, Zapier, n8n — for connecting AI to other applications. The trigger → AI decision → action chain automates everyday processes.
Safety and monitoring
Check that the agent is moving in the right direction at each step. The instruction 'if unsure — stop and ask' is critically important.
Weak vs strong prompt
Analyze my emails and pick out the important ones.
Goal: sort today's emails by importance. Steps: 1) read each email's subject and sender; 2) categorize as urgent/important/routine; 3) list urgent emails separately; 4) for each, add a recommended action. If categorization is unclear — stop and ask me. Format: table.
The second prompt sets clear steps, categories, and a stop condition. The agent knows what to do and doesn't make decisions on its own in ambiguous situations.
Ready prompt template
Copy and adaptGoal: [final result]. Available tools: [list of tools]. Steps: 1) [first subtask]; 2) [second subtask]; 3) [third subtask]. After each step, pass the result to the next. If at any step you're unsure — stop, notify me, and wait for instructions. Final output format: [format].
Why it works
The agent calls the necessary tool at each step: search, computation, file reading, API request — this fundamentally differs from regular chat.
If steps aren't clearly written, the agent will guess. Specific formulation of each subtask reduces error probability.
The 'if unsure — stop' instruction prevents the agent from going far in the wrong direction.
Make.com/Zapier + AI: trigger → AI analysis → action. Accessible and affordable automation for small business.
Practice
- Identify one repeating task in your work: for example, writing a weekly report.
- Break it into steps: data collection, analysis, writing, sending.
- Convert each step into a subtask for the agent and write an instruction.
- Test in ChatGPT or Claude and identify which steps require human involvement.
Mini-project
Mini-project: automating one workflow
Choose a repeating task and automate it with an agentic prompt or a workflow in Make.com/Zapier.
Tasks
- Identify the repeating task and describe its steps.
- Convert each step into a subtask for the agent.
- Test in ChatGPT or Claude.
- Identify steps that require human involvement.
Deliverables
- 1 agentic prompt or workflow diagram
- test results
- list of steps requiring human oversight
Checklist
Common mistakes
- not writing instructions for the agent and expecting it to 'figure it out'
- not adding stop points — the agent can build a long chain of errors
- not checking each step — finding errors after completion is harder
- immediately writing complex interconnected tasks for the agent — better to start simple
Lesson FAQ
Do you need programming skills to work with agentic AI?
No, not for basic use. In ChatGPT, Claude, and platforms like Make.com/Zapier, you can build agentic workflows without writing code.
What happens if the agent makes a mistake?
The agent continues building the chain based on the error — that's why monitoring is important. A checkpoint at every important step reduces this risk.