What exactly is agentic AI?

While a classic chatbot like ChatGPT answers a single query with a single response, agentic AI can independently plan multiple steps, make decisions, use tools, and iteratively work toward a goal. Imagine an assistant you tell "book me a flight to London for next Tuesday, the cheapest possible, and send the itinerary to my boss" — and it does the whole thing.

The key difference from regular LLMs is the feedback loop: the agent plans steps → executes an action → evaluates the result → adjusts the plan → continues further. This cycle repeats until the goal is achieved, or until the agent hits an insurmountable obstacle.

In 2025 and 2026, we're witnessing an explosion of agentic systems. OpenAI launched the Responses API and Agents SDK, Google has Project Mariner for browser control, Anthropic offers tool use in Claude, and the open-source scene is churning out frameworks at a pace that's hard to keep up with.

Why "build" and not just "read"?

Professor Sébastien Martin from the Kellogg School of Management, who teaches agentic systems as part of the MBA program, emphasizes that passive consumption of AI information isn't enough. "Students who actually build an agent understand the technology's limits much faster than those who just listen to lectures about it," he notes in his teaching.

The hands-on approach has three advantages:

1. Immediate feedback. When an agent fails (and it will fail often), you see right away why — tool hallucination, poor decision-making in the loop, missing context. Theory becomes reality within minutes.

2. Understanding architecture. Building an agent from scratch reveals how orchestration logic, memory management, tool calling, and error handling work. These are concepts you won't pick up from a press release.

3. Distinguishing hype from reality. Most demo videos show agentic systems in ideal conditions. When you build them yourself, you'll discover that the real success rate of autonomous tasks is significantly lower — according to the SWE-bench benchmark, the best agents achieve around 50–70% success on complex coding tasks.

Where to start: framework overview

For beginners and experienced developers alike, there are several paths today:

LangChain and LangGraph

The most widespread open-source framework for building LLM applications. LangGraph adds the ability to define agents as state graphs — ideal for complex workflows. Price: open-source free, paid LangSmith version for monitoring from $39/month. Czech availability: fully available, documentation in English.

Microsoft AutoGen

A framework from Microsoft Research focused on multi-agent conversations — multiple agents communicate with each other and solve tasks together. AutoGen 0.4 brought a fundamental redesign with asynchronous architecture in January 2026. Price: open-source free, you only pay for API calls to models.

CrewAI

Built on LangChain, but focused purely on multi-agent systems with roles (you define a "researcher," "editor," "fact-checker"). Price: open-source free, CrewAI Enterprise from $10/user/month. Great for rapid prototyping.

OpenAI Agents SDK

The official SDK from OpenAI, formerly known as Swarm. Offers the lowest entry barrier if you already use the OpenAI API. Agents pass context between each other (handoffs) and use tools via function calling. Price: SDK free, you only pay for API tokens.

Anthropic Tool Use and MCP

Anthropic is paving the way through the Model Context Protocol (MCP) — an open standard for connecting agents to external data and tools. Instead of custom orchestration, agents rely on the capabilities of the Claude model. Price: SDK and MCP free, API tokens according to the plan (Claude Haiku from $0.25/million tokens).

Your first agent in 30 minutes

We'll build a simple agent using the OpenAI Agents SDK that can search for current information and answer questions. Python code:

Step 1: Install the SDK — pip install openai-agents

Step 2: Define a tool (a function for searching) and the agent that will use it.

Step 3: Run the agent with a query and watch how it plans steps, calls the tool, and returns results.

The first run usually reveals surprising behavior — the agent may call the tool repeatedly, interpret results incorrectly, or get stuck in a loop. These moments are exactly the most valuable lesson: you'll discover that orchestration logic (when to stop, how to evaluate results) is much more important than the model itself.

Complete tutorials for individual frameworks can be found on their official websites — LangChain offers interactive LangGraph tutorials, OpenAI has clear Agents quickstart, and CrewAI provides a step-by-step guide.

What you'll learn by building an agent

Tool calling isn't magic. You'll discover that the tool definition (its description, parameters, return value) fundamentally affects how well the agent uses the tool. A poorly described tool = the agent ignores it or uses it incorrectly.

Context window is precious. Every agent step consumes tokens. An agent that performs 15 iterations with full context can cost $2–5 in API fees for a single task. Memory and context optimization is a key skill.

Evaluation is harder than building. How do you know if the agent completed the task correctly? Unlike classic software where you test deterministic output, agentic systems require a new approach to testing — from LLM-as-a-judge to specialized eval platforms like Braintrust.

Security is not optional. An agent with access to APIs, databases, or the file system must have a sandbox and limited permissions. Prompt injection attacks, where a user convinces the agent to perform an unwanted action, are a real threat.

Agent operation costs: realistic numbers

Let's look at concrete numbers. A simple agent calling GPT-4o (or Claude 4 Sonnet) consumes approximately 5,000–15,000 input tokens and 500–1,500 output tokens per task of 3–5 steps. At current prices (July 2026), that comes out to:

• GPT-4o: $2.50/1M input, $10/1M output → approx. $0.02–0.05 per task
• Claude 4 Sonnet: $3/1M input, $15/1M output → approx. $0.03–0.07 per task
• Gemini 2.5 Pro: $1.25/1M input (up to 200K tokens) → approx. $0.01–0.03 per task
• Open-source models (Llama 4, DeepSeek V3): your own GPU/server, from $0.50/hour on RunPod

For experimentation and learning, I recommend starting with GPT-4o mini ($0.15/1M input) or Claude Haiku ($0.25/1M input) — learning will cost you literally pennies.

The Czech and European footprint

In the Czech Republic, agentic AI is still in its early stages, but the first signs are already here. Czech companies like Apify (web scraping agents), Keboola (data agents), or Rossum (document agents) are integrating agentic elements into their products. For developers, English proficiency is essential — all documentation, tutorials, and community forums are in English. Czech as a language for prompt engineering with agents works, but for more complex tasks I recommend English — models still have a higher error rate with Czech instructions.

European regulation through the EU AI Act applies to agentic systems from August 2026 — agents making decisions with legal or financial impacts fall into the "high risk" category and require human oversight. For companies in the EU, this means that fully autonomous agentic systems without the possibility of human control will not be in compliance with the regulation.

Where agentic AI is today: successes and limits

Agentic systems excel in tasks with clear structure: customer support (agent searches the knowledge base, finds relevant articles, suggests solutions), code review and debugging (agent analyzes code, runs tests, proposes fixes), research assistants (agent searches multiple sources, synthesizes information, creates a report).

On the other hand, they fail in tasks requiring long-term planning (more than 20 steps), ambiguous goals ("do it as well as possible"), or working with unstructured visual data (controlling GUI applications).

According to the latest data from the GAIA (General AI Assistant benchmark), the best agentic systems achieve a success rate of around 50–60% — compared to 92% human success. It's a huge leap from 2024, when agents were around 15%, but there's still a long road ahead.