AI & Machine LearningIntermediate
Agentic AI
Artificial intelligence systems that can act autonomously, make decisions, and execute complex tasks with minimal human intervention.
Pronunciation
/eɪˈdʒɛntɪk eɪ aɪ/
"ay-JEN-tik ay-eye"
Listen on:Forvo
What is it
Agentic AI refers to artificial intelligence systems that can:
- Act autonomously without step-by-step instructions
- Make decisions based on context and goals
- Use tools to complete tasks
- Iterate until achieving the desired result
Unlike a traditional chatbot that only answers questions, an AI agent does things.
Common pronunciation mistakes
- ❌ “AY-jen-tik” (stressing the first syllable, as in “agent”; the stress is on “JEN”)
- ❌ “a-jen-TIK” (stressing the last syllable)
Chatbot vs AI Agent
| Aspect | Traditional Chatbot | AI Agent |
|---|---|---|
| Interaction | Question → Answer | Goal → Actions → Result |
| Autonomy | Waits for instructions | Decides what to do |
| Tools | Doesn’t use tools | Uses APIs, files, code |
| Iteration | One response | Multiple attempts |
| Context | Current conversation | Entire project |
How an AI Agent Works
┌─────────────────────────────────────────────────────────┐
│ AGENT CYCLE │
├─────────────────────────────────────────────────────────┤
│ │
│ User: "Implement JWT authentication in my API" │
│ │ │
│ ▼ │
│ ┌─────────────────────────────────────────────────┐ │
│ │ 1. PLAN │ │
│ │ - Analyze project structure │ │
│ │ - Identify files to modify │ │
│ │ - Determine required dependencies │ │
│ └─────────────────────────────────────────────────┘ │
│ │ │
│ ▼ │
│ ┌─────────────────────────────────────────────────┐ │
│ │ 2. EXECUTE │ │
│ │ - Install jsonwebtoken │ │
│ │ - Create authentication middleware │ │
│ │ - Modify protected routes │ │
│ └─────────────────────────────────────────────────┘ │
│ │ │
│ ▼ │
│ ┌─────────────────────────────────────────────────┐ │
│ │ 3. VERIFY │ │
│ │ - Run tests │ │
│ │ - Review errors │ │
│ │ - Iterate if necessary │ │
│ └─────────────────────────────────────────────────┘ │
│ │ │
│ ▼ │
│ Result: "Authentication implemented. Here's the │
│ usage documentation..." │
│ │
└─────────────────────────────────────────────────────────┘Examples of AI Agents
For software development
| Agent | Description | Use case |
|---|---|---|
| Claude Code | Anthropic’s CLI agent | Full-stack development |
| GitHub Copilot coding agent | Agent in GitHub | Assigned issues → draft PRs |
| Devin (Cognition) | Autonomous agent | Complex engineering tasks |
| Cursor Agent | Inside the IDE | Multi-file refactoring |
For productivity
| Agent | Description | Use case |
|---|---|---|
| Computer Use (Anthropic) | Computer control | Task automation in graphical interfaces |
| Browser agents | Web navigation | Reservations, purchases, forms |
| Office-suite assistants (Gemini, Microsoft Copilot) | Built into email and documents | Emails, calendar |
Real-World Example: Agent vs Chatbot
Task: “Fix the shopping cart bug”
With traditional Chatbot:
User: There's a bug in the cart
AI: Can you describe the bug?
User: The total doesn't update
AI: You probably need to recalculate the total after removing items.
Here's a code example...
User: Where do I put it?
AI: In your remove items function...
[10 more messages]With AI Agent:
User: There's a bug in the cart - the total doesn't update when removing items
Agent: I'll investigate and fix it.
[Agent automatically:]
1. Reads cart files
2. Identifies removeItem function
3. Finds missing total recalculation
4. Modifies the code
5. Runs tests
6. Creates commit
Agent: Fixed. The problem was in cart.service.ts line 45.
It forgot to call recalculateTotal() after splice().
I already committed the fix.Components of an Agentic System
┌─────────────────────────────────────────────────────────┐
│ AGENTIC ARCHITECTURE │
├─────────────────────────────────────────────────────────┤
│ │
│ ┌──────────────┐ │
│ │ Base LLM │ Agent's brain (Claude, GPT) │
│ └──────┬───────┘ │
│ │ │
│ ┌──────┴───────┐ │
│ │ Memory │ Context, history, state │
│ └──────┬───────┘ │
│ │ │
│ ┌──────┴───────┐ │
│ │ Tools │ APIs, files, code, web │
│ └──────┬───────┘ │
│ │ │
│ ┌──────┴───────┐ │
│ │ Planner │ Decomposes tasks, decides steps │
│ └──────────────┘ │
│ │
└─────────────────────────────────────────────────────────┘Risks and Considerations
Risks
| Risk | Mitigation |
|---|---|
| Unwanted actions | Granular permissions, human confirmation |
| Infinite loops | Iteration limits, timeouts |
| Excessive costs | Token limits, monitoring |
| Security | Sandboxing, principle of least privilege |
When NOT to use agents
- Simple single-response tasks
- When total step-by-step control is required
- In critical systems without supervision
- Without established security policies
Where it’s heading
“AI agents will proliferate in 2026 and play a bigger role in daily work, acting more like teammates than tools.” — Vasu Jakkal, corporate vice president of Microsoft Security, in What’s next in AI: 7 trends to watch in 2026 (Microsoft, December 2025)
This is an industry prediction, not an established fact: real-world adoption varies widely by sector, task risk, and the maturity of security controls.
Related terms
- MCP - Protocol to connect agents to services
- LLM - Large Language Model, the agent’s brain
- Prompt Engineering - How to instruct agents
Remember: AI agents are powerful but require supervision. Start with low-risk tasks and gradually increase autonomy as you gain confidence.