# Agentic AI

What AI agents are, how they work, and every protocol, framework, pattern and operating practice on this site, in one map.

## Overview

Agentic AI describes AI systems that do not just answer a question but work toward a goal. An agent receives a task, decides which steps to take, uses tools such as search, databases, code or other software, looks at the results and adjusts until the task is done or it needs a person's input.

The difference from a plain chatbot is like the difference between asking a librarian where a book is and asking a research assistant to write you a briefing. The librarian answers one question. The research assistant plans the work, pulls sources, reads them, notices gaps, looks again and hands you a finished result. Agentic AI is the research assistant: a language model placed inside a loop with tools, memory and rules.

This page is the map for everything about agents on this site. It explains the core ideas briefly and links to a detailed page for each protocol, framework, pattern and operating practice.

## What makes a system agentic

Most agents share five parts.

| Part | Role | Read more |
|---|---|---|
| Model | Reasons about the goal and chooses the next step | [ReAct loop](/tools/react-loop) |
| Tools | Let the model read and change the outside world | [Function calling](/tools/function-calling), [MCP](/tools/mcp) |
| Memory | Keeps context within a task and across sessions | [Agent memory](/tools/agent-memory) |
| Orchestration | Controls the loop, branching, retries and handoffs | [LangGraph](/tools/langgraph), [Multi-agent systems](/tools/multi-agent-systems) |
| Controls | Keep behaviour safe, measurable and affordable | [Guardrails](/tools/guardrails), [LLM-as-judge](/tools/llm-as-judge), [Prompt-injection defense](/tools/prompt-injection-defense) |

Agency is a spectrum, not a switch. At one end sits a fixed workflow in which a model fills in one step. At the other sits an autonomous agent that chooses its own steps for hours. Most useful production systems sit in between: the model decides within clear boundaries, and code enforces the rules.

## The protocols that connect agents

Two open standards now shape how agents connect to the world. **MCP** connects an agent to tools, data and prompts through servers that any MCP-capable application can use. **A2A** connects one agent to another, so agents from different vendors can discover each other, delegate tasks and exchange results. Both are stewarded by the Agentic AI Foundation under the Linux Foundation. Underneath both sits **function calling**, the way a model expresses that it wants a tool.

- [MCP: the Model Context Protocol](/tools/mcp)
- [A2A: the Agent2Agent protocol](/tools/a2a)
- [Function calling](/tools/function-calling)

## Patterns for how agents think and work

- [ReAct loop](/tools/react-loop): think, act, observe, repeat. The default loop of most agents.
- [Multi-agent systems](/tools/multi-agent-systems): orchestrators, workers, routers, handoffs and debate, and when splitting work pays off.
- [LLM-as-judge](/tools/llm-as-judge): using a model to grade outputs and trajectories against a rubric.
- [Guardrails](/tools/guardrails): layered checks on input, context, actions and output.
- [Agentic RAG](/tools/agentic-rag): retrieval driven by an agent that plans, checks evidence and searches again.

## Running agents in production

Agents fail in ways ordinary software does not. They can pick the wrong tool, loop, follow injected instructions or quietly produce plausible but wrong results. Production agents therefore need the same discipline as any critical system, plus a few extra habits:

1. **Evaluate before and after every change.** Keep a set of real tasks with expected outcomes, measure success, cost and steps, and use calibrated judges for open-ended quality.
2. **Trace every run.** Record each model call, tool call and handoff under one trace id.
3. **Limit power.** Give tools the least privilege they need, require approval for side effects, and cap steps, time and spend.
4. **Treat outside content as hostile.** Web pages, emails, documents and tool results can carry injected instructions.
5. **Prefer simple designs.** Start with one agent and few tools; add agents, memory and autonomy only when evaluation shows a need.

## The full map

Every item in the agentic AI part of my stack, grouped as on the [tools page](/tools#agents). Linked items open a full guide.

### Agent SDKs

- [Claude Agent SDK](https://www.ozercemkelahmet.com/tools/claude-agent-sdk.md)
- [OpenAI Agents SDK](https://www.ozercemkelahmet.com/tools/openai-agents-sdk.md)
- [Google ADK](https://www.ozercemkelahmet.com/tools/google-adk.md)
- [Microsoft Agent Framework](https://www.ozercemkelahmet.com/tools/microsoft-agent-framework.md)

### Frontier topics

- [Agent harness](https://www.ozercemkelahmet.com/tools/agent-harness.md)
- [Agent identity and delegation](https://www.ozercemkelahmet.com/tools/agent-identity-and-delegation.md)
- [Agentic commerce](https://www.ozercemkelahmet.com/tools/agentic-commerce.md)
- [Generative UI](https://www.ozercemkelahmet.com/tools/generative-ui.md)

### Frameworks

- [LangGraph](https://www.ozercemkelahmet.com/tools/langgraph.md)
- [CrewAI](https://www.ozercemkelahmet.com/tools/crewai.md)
- [Pydantic AI](https://www.ozercemkelahmet.com/tools/pydantic-ai.md)
- [Mastra](https://www.ozercemkelahmet.com/tools/mastra.md)
- [smolagents](https://www.ozercemkelahmet.com/tools/smolagents.md)
- [LlamaIndex](https://www.ozercemkelahmet.com/tools/llamaindex.md)
- [DSPy](https://www.ozercemkelahmet.com/tools/dspy.md)

### Protocols

- [MCP](https://www.ozercemkelahmet.com/tools/mcp.md)
- [A2A](https://www.ozercemkelahmet.com/tools/a2a.md)
- [Function calling](https://www.ozercemkelahmet.com/tools/function-calling.md)
- [Structured outputs](https://www.ozercemkelahmet.com/tools/structured-outputs.md)
- [OAuth for tool access](https://www.ozercemkelahmet.com/tools/oauth-for-tool-access.md)

### Patterns

- [ReAct loop](https://www.ozercemkelahmet.com/tools/react-loop.md)
- [Plan-and-execute](https://www.ozercemkelahmet.com/tools/plan-and-execute.md)
- [Reflection](https://www.ozercemkelahmet.com/tools/reflection.md)
- [Tree-of-thoughts](https://www.ozercemkelahmet.com/tools/tree-of-thoughts.md)
- [Orchestrator-workers](https://www.ozercemkelahmet.com/tools/orchestrator-workers.md)
- [Multi-agent systems](https://www.ozercemkelahmet.com/tools/multi-agent-systems.md)
- [Human-in-the-loop](https://www.ozercemkelahmet.com/tools/human-in-the-loop.md)
- [LLM-as-judge](https://www.ozercemkelahmet.com/tools/llm-as-judge.md)
- [Guardrails](https://www.ozercemkelahmet.com/tools/guardrails.md)

### Coding and computer agents

- [Claude Code](https://www.ozercemkelahmet.com/tools/claude-code.md)
- [Computer use](https://www.ozercemkelahmet.com/tools/computer-use.md)
- [Browser agents](https://www.ozercemkelahmet.com/tools/browser-agents.md)
- [Long-running tasks](https://www.ozercemkelahmet.com/tools/long-running-tasks.md)
- [Agent memory](https://www.ozercemkelahmet.com/tools/agent-memory.md)
- [Skills](https://www.ozercemkelahmet.com/tools/skills.md)
- [Sub-agents](https://www.ozercemkelahmet.com/tools/sub-agents.md)

### Workflow automation

- [n8n](https://www.ozercemkelahmet.com/tools/n8n.md)
- [Temporal](https://www.ozercemkelahmet.com/tools/temporal.md)
- [Camunda](https://www.ozercemkelahmet.com/tools/camunda.md)
- [Apache Airflow](https://www.ozercemkelahmet.com/tools/apache-airflow.md)
- [UiPath](https://www.ozercemkelahmet.com/tools/uipath.md)

### Operations

- [Tracing](https://www.ozercemkelahmet.com/tools/tracing.md)
- [Evals](https://www.ozercemkelahmet.com/tools/evals.md)
- [Regression boards](https://www.ozercemkelahmet.com/tools/regression-boards.md)
- [Cost and latency budgets](https://www.ozercemkelahmet.com/tools/cost-and-latency-budgets.md)
- [Sandboxing](https://www.ozercemkelahmet.com/tools/sandboxing.md)
- [Prompt-injection defense](https://www.ozercemkelahmet.com/tools/prompt-injection-defense.md)

## Related pages on this site

- [Agentic RAG](/tools/agentic-rag) and the article [Retrieval-augmented generation, from naive to agentic](/articles/rag-from-naive-to-agentic)
- [Multi-agent reinforcement learning](/tools/multi-agent-rl) and [Constitutional AI](/tools/constitutional-ai)
- Evaluation tools: [DeepEval](/tools/deepeval), [RAGAS](/tools/ragas), [TruLens](/tools/trulens)
- Connect your own agent to this site: the read-only MCP server at `https://mcp.ozercemkelahmet.com/mcp` offers `search`, `read_page` and `list_topics` over every page, with no sign-in.
- Using a tabular foundation model from an agent: [Agents and LLMs](/model/integrations/agents) and the [MCP server](/model/mcp)

## Frequently asked questions

### What is agentic AI?

Agentic AI refers to AI systems that pursue a goal by planning steps, using tools, observing results and adjusting, rather than producing a single answer to a single prompt.

### What is the difference between an AI agent and a chatbot?

A chatbot answers messages. An agent works toward a goal: it decides which actions to take, calls tools, reads results and continues until the task is complete or it needs human input.

### What are MCP and A2A?

MCP is an open protocol that connects agents to tools and data. A2A is an open protocol that connects agents to other agents. Both are stewarded by the Agentic AI Foundation.

### Do I need a multi-agent system?

Usually not at first. A single agent with a few well-designed tools solves most problems. Add agents when tasks split into parallel parts, need very different tools, or benefit from an independent reviewer.

### How do you make AI agents safe?

Combine least-privilege tools, human approval for side effects, budgets, guardrails on input and output, defences against prompt injection, full tracing and regular evaluation.

## Summary

Agentic AI places a language model inside a loop with tools, memory, orchestration and controls, so it can pursue goals instead of answering single prompts. MCP and A2A standardise how agents reach tools and each other; patterns such as ReAct, multi-agent orchestration and LLM-as-judge shape how they work; guardrails, injection defences, tracing and evaluation keep them safe and reliable. Use the map above to go deeper into any part.