Agentic AI: planning, reasoning, and acting with Autonomy

What is agentic AI?#

Agentic AI refers to systems that don’t just answer questions, but can plan, decide, and act on their own.

They use LLMs + tools + memory to handle multi-step tasks without being told exactly what to do at each step.

For example:#

• You say “Book me a trip to Delhi this weekend under Rs.5000.”

  • An agentic AI might:

    • Search for flights and buses.
    • Check hotels.
    • Compare costs.
    • Make bookings via APIs.
    • Email you the plan.

  • This is way beyond generating text: this is actual decision + action.

---

Characteristics of agentic AI#

1. Goal-oriented planning

   • It starts with a goal, then breaks it into smaller steps and executes.

2. Multi-step reasoning

   • It doesn’t stop at one reply: it iteratively reasons until the goal is met.

3. Autonomous decision-making

   • It chooses the best path or tool on its own, based on whatever context it has.

4. Uses memory, tools, and knowledge

   • Remembers past events (long-term or conversational memory), uses APIs and scripts (tools), and refers to docs or rules (knowledge base).

---

What makes agentic AI different from basic LLMs?#

• Normal LLMs (ChatGPT, Gemini, Perplexity, etc.) are passive:
  • They just respond to a prompt.
• Agentic systems are active:
  • They observe, think, plan, act, and even revise their actions.

---

Basic Agentic Structure#

---

Agentic AI vs AI workflow#

Term	What it means
AI workflow	Pre-defined steps using AI (e.g. RAG pipeline)
Agent	Dynamic reasoning & acting system that adapts
ReAct Agent	Reason + Act: Think, Decide, take Action
Control logic	Like if/else to guide actions
Tools	External Functions/APIs
Memory	Past info/context retained
Planner	Creates a step-wise plan

---

LLMs in this context#

• LLMs (like GPT, Gemini, Claude) are the core brains of the system.
• But on their own, they can’t remember, act, or plan.
• We wrap them in agents to extend their ability.

---

Agentic AI examples#

1. AI coding assistant
   • Understands a bug report, edits code, runs tests, commits.
2. Travel agent
   • The trip-booking example above. There are a gazillion use cases like it across domains.
3. A few more: an HR onboarding bot, an agent that posts for you on LinkedIn, the list goes on~

---

Let’s recap#

• LLM = brain
• Agent = brain + body (tools, memory, planner)
• Agentic AI = an autonomous worker that thinks and achieves any goal given to it
• Multi-turn planning = not just simple answers, but strategic answers driven by context
• ReAct loop = reason -> act -> observe -> repeat

Important: before you go, read this excellent post by Anthropic on building effective agents.