وكلاء الذكاء الاصطناعي: مدخل عملي إلى Agentic AIAI Agents: A Practical Introduction to Agentic AI

Key terms

Agentic AI

AI systems that do more than answer a question. They plan and carry out a sequence of steps to reach a goal.

Agent

A program that uses a model such as Claude to decide the next step, calls tools to carry it out, checks the result and continues.

Tools

The capabilities you give the agent: reading email, searching the web, writing to a sheet, running a command. An agent is only as capable as its tools.

MCP

An open standard for connecting a model to tools and data sources in one consistent way, instead of building a custom connection for each tool.

Human in the loop

The agent prepares the work and waits for your approval before any important step, such as sending, paying or deleting.

The steps

1. 1

   Step 1

   See the difference between a chat and an agent

   In a normal chat you ask, the model answers and you do the work.

   With an agent you give it a goal. It breaks the goal into steps, calls tools to carry them out, checks the result and keeps going until it is done or needs you.

   That loop of plan, act, check is the core of agentic AI. Everything else is detail.

   PromptTest the idea in a chat

   Copy

   I want to know if this task suits an agent or if a normal chat is enough:
   [THE TASK AS IT HAPPENS TODAY]
   
   Break it into steps. For each step: is it a decision or an action? What tool or data does it need?
   Then tell me plainly: agent or chat? And why?

2. 2

   Step 2

   Pick one repeated task

   Do not start with an agent that does everything. Start with one task that repeats every day or every week, with a clear input and a clear result.

   Examples: triaging the inbox, preparing a weekly report from a spreadsheet, answering repeated customer questions.

   A good first task is boring and repetitive, and a mistake in it is easy to fix.

   Tip: If a mistake in the task is costly or cannot be undone, save it for after your first agent.

3. 3

   Step 3

   Map the tools and data

   List everything the agent needs to read or do: email, calendar, files, a spreadsheet, a website.

   For each tool, set the permission: read only, or read and write. Start with the least access that works.

   If you use Claude, many of these tools are available through MCP, so you connect them instead of building them from scratch. Test each tool on its own before wiring them together.

   PromptTool map

   Copy

   Here is a task I want to build an agent for: [TASK].
   These are the tools and data I have: [EMAIL, CALENDAR, GOOGLE DRIVE, SPREADSHEET, OTHERS].
   
   Make me a table: the tool, why the agent needs it, the least access that is enough (read or write), and the risk if it misuses it.

4. 4

   Step 4

   Write the agent instructions

   The instructions are the contract between you and the agent. The agent does not understand your intent. It follows what you wrote.

   Include the mission in one sentence, the scope, what is forbidden, the allowed tools, when to stop and what to do in an unexpected situation.

   The template below is a good starting point.

   PromptAgent instructions template

   Copy

   You are a [TASK TYPE] agent for [COMPANY OR PROJECT].
   Your single mission: [ONE SENTENCE].
   
   Scope: work only on [DATA, FOLDER OR CHANNEL].
   Forbidden: [WHAT YOU MUST NOT DO].
   Allowed tools: [TOOLS].
   
   Stop when: [CLEAR COMPLETION CONDITION].
   If you hit a situation these instructions do not cover, stop and tell me instead of improvising.
   
   Do not send, delete or pay for anything before I approve.
   At the end, list everything you did and why.

5. 5

   Step 5

   Start in draft and approve mode

   For the first weeks, the agent drafts and you approve. Every message, every reply, every change waits for your sign off.

   This protects you from mistakes and shows you where the agent goes wrong, so you can fix its instructions.

   Trust is earned through runs. Do not give it full autonomy on day one.

   Tip: Even once you trust it, keep human approval on anything sensitive or irreversible.

6. 6

   Step 6

   Test on real cases and log decisions

   Run the agent on small cases from your real work, not perfect examples.

   Ask it to write what it is about to do and why before every step. That log shows you where it drifts and makes review faster.

   Every mistake that repeats becomes a written rule in the instructions.

   PromptRun review

   Copy

   Here is the agent's run log and its result:
   [PASTE THE LOG AND RESULT]
   
   Identify: where did it leave the mission or the scope? Where did it guess instead of asking? Which tool did it use without need?
   Then suggest specific edits to its instructions that prevent these mistakes, without rewriting the whole thing.

7. 7

   Step 7

   Expand step by step

   Once the agent is reliable on its first task, add a second, related task, or relax approval on routine steps only.

   Every repeated task can become its own skill with its own instructions and templates.

   A successful agent is built one step at a time, not all at once.

Ready prompts to copy

Here is a list of tasks I repeat at work:
[TASKS]

Rank them by how well they fit a first agent. Criteria: they repeat often, the input and result are clear, mistakes are easy to fix, and they need few tools.
Explain your ranking in one line per task.