Agents & MCP: when AI stops talking and starts acting
tier II · practitioner · 10 min · interactive article ✦
Until now, your AI assistant lived in a chat window: it answered, you copy-pasted. The next generation acts: it consults your tools, chains steps, executes tasks, and reports back. Two words frame this shift: agent and MCP. And one question sums it up: who sets the limits?
The agent: a digital colleague with a mission
An AI agent is a system built around an LLM, given a goal rather than a question: “prepare the weekly project review.” To get there, it breaks the task down (check tickets, spot blockers, write), uses tools (read a dashboard, query a database, send a draft) and loops until the result. The difference with chat: you no longer drive each step — you set the mission and the limits.
MCP: the universal socket
To act, the agent needs access to your tools — calendar, CRM, project management, documents. Historically, each connection was a custom build: costly, fragile, unauditable. MCP (Model Context Protocol) standardizes that connection: a vendor exposes an “MCP server” for its tool, and any compatible assistant can plug into it — with defined permissions.
one socket, permissions per tool — instead of a custom connector per AI and per tool
Think of USB: before, every device had its proprietary cable; after, one socket. MCP is becoming the USB of enterprise AI.
The question is no longer “can we?” but “how do we govern it?”
Scope: which tools does the agent access, read or write?
Guardrails: which actions require human validation? (sending an email ≠ preparing a draft of it)
Traceability: is every agent action logged? Who audits it?
Reversibility: if the agent gets it wrong, can it be undone — and who is responsible?
This is exactly where trained managers make the difference: technology makes action possible, your role is to make it governed. An agent with no guardrails isn't innovation, it's an operational risk with a nice interface.
test yourself — just like in the path
Your team wants to give an AI agent write access to the project-management tool. Your manager's reflex?
✓ Neither naivety nor blockage: governed access. It's the stance that separates a trained manager from a mere user — and it's the whole point of tier II.
Between recklessness and outright refusal lies governance. Try again.
Going deeper
Agent, assistant, copilot… who does what?
The assistant answers when asked. The copilot suggests continuously inside a given tool. The agent receives a goal and chains the steps itself to reach it. Autonomy rises — and with it, the need for governance.
Where to start without risk?
With a read-only agent on a non-sensitive scope: weekly ticket summary, monitoring a document base, preparing drafts. No writes, systematic human validation on output. You learn the mechanics without exposing operations.
The question to ask your IT department
“Do our tools already expose MCP servers, and if so, with what default permissions?” Most major vendors have published them since 2025 — the socket likely already exists in your stack, plugged in or not.
Frequently asked questions
What's the difference between a chatbot and an AI agent?
A chatbot answers; an agent pursues a goal: it breaks down the task, uses tools and loops until the result — under human supervision at the key points.
What is MCP (Model Context Protocol)?
An open standard that connects AI to your tools and data (calendar, CRM, files), with defined permissions. It's the “universal socket” between AI and your systems.
Does an AI agent act with no control?
No. At critical points, human validation remains the rule: you set a clear goal and guardrails. Autonomy is bounded.
Does MCP make AI more intelligent?
No. MCP connects AI to your tools; it doesn't increase the model's intelligence. It's plumbing, not compute power.
This notion is a full level of the IAPLC path.
In the course, you practice it on YOUR context: lessons generated for your job, interactive exercises, gamified progression.