← AI Terminology
Agentic AI
Agentic AI refers to AI systems that operate autonomously over extended tasks — perceiving context, planning steps, taking actions, and iterating — with minimal human intervention per action.
It marks the transition from AI as a single-turn tool to AI as a persistent, goal-directed actor.
It marks the transition from AI as a single-turn tool to AI as a persistent, goal-directed actor.
Why It Matters in AI
Agentic AI is the architectural leap that turns LLMs into software workers: writing code, browsing the internet, managing files, calling APIs, and revising outputs based on results. It dramatically expands what AI can automate but also introduces new risks — agents can take consequential, hard-to-reverse real-world actions. Understanding agentic AI is essential for evaluating both the opportunity and the safety requirements of frontier AI systems.
Key Points
| Aspect | Description |
|---|---|
| Tools | Web search, code execution, file I/O, API calls, computer use, database queries |
| Memory | Context window (short-term), vector stores or files (long-term), episodic logs |
| Core loop | Observe → Plan → Act → Observe (repeat) — often called the ReAct or agent loop |
| Key products | Devin, Operator (OpenAI), Claude computer use, GitHub Copilot Workspace, AutoGen |
| Risk factors | Irreversible actions, prompt injection, goal misgeneralisation, resource acquisition |
| Autonomy levels | Single-step assistant → multi-step agent → fully autonomous agent (increasing risk and capability) |
Simple Analogy
A vending machine is transactional AI — you press a button, you get a result. Agentic AI is more like hiring a contractor: you describe the goal, they plan the work, execute multiple steps, deal with obstacles, and deliver the outcome — checking in with you only when genuinely stuck.
Common Usage Examples
- Claude instructed to "research competitors, build a comparison table, and email it" — executes all steps autonomously
- GitHub Copilot Workspace plans and implements entire feature branches from a natural-language issue
- AutoGen: orchestrates a team of specialised sub-agents (planner, coder, reviewer) on a shared goal
- Agentic AI in customer support: resolves tickets end-to-end including looking up accounts and issuing refunds
- Safety measure: "minimal footprint" principle — agents should request only permissions needed for the immediate task
Summary
In short: Agentic AI is the shift from AI that answers to AI that acts — and the central design challenge is giving it enough autonomy to be useful while keeping humans in meaningful control.