The world of software engineering is undergoing a fundamental transformation. It's no longer just about AI helping to complete lines of code (as with GitHub Copilot). The new wave, referred to as Agentic AI, represents systems that possess a certain degree of autonomy. They can define steps to achieve a goal, use tools, correct their own errors, and in many cases, independently manage entire projects.
From Chatbot to Autonomous Agent: What Exactly is Agentic AI?
To understand the current trend, it is necessary to distinguish between generative AI, which we know from ChatGPT or Claude, and Agentic AI. While standard models are reactive – they answer your question and wait for the next one – agentic systems are proactive. They have the ability to plan and execute.
If you write to a classic LLM: "Write me a function for email validation," you will get code. However, if you assign a task to Agentic AI, such as Devin, you can tell it: "Find all outdated email validation methods in our repository, replace them with a new implementation, and ensure all tests still pass." The agent will break the task into subtasks, go through files, make changes, and in case of an error, attempt a fix before presenting you with the final report.
Technology Comparison: LLM vs. AI Agents
Within the current market, we can see a clear division:
- Standard Models (GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro): Excellent for generating text and code fragments, but require constant human supervision and iteration.
- AI Programming Agents (Devin, OpenDevin): Specialized systems built on these models that have access to a terminal, browser, and file system. Their goal is to minimize human intervention in repetitive tasks.
The Nubank Case: When AI Handles Millions of Lines of Code
One of the most striking proofs of Agentic AI's effectiveness is its successful implementation at the digital bank Nubank. The company faced enormous technical debt – a monolithic ETL system containing over 6 million lines of code that required refactoring. The traditional path would have required thousands of engineers and years of work.
Thanks to the deployment of the Devin agent, Nubank managed to achieve a 12-fold increase in efficiency in engineering hours and a 20-fold cost saving. What would have taken months or years was accomplished in a matter of weeks thanks to the autonomous work of the agents. This example shows that Agentic AI is not just for small scripts, but for the critical infrastructure of the largest financial institutions.
Recruitment Boom: Why Are Companies Hiring More People When AI Codes?
It may seem paradoxical that while AI automates programming, reports like the one from InfotechLead speak of growing recruitment activities. The key, however, is a change in the worker's profile. Companies in the finance and pharmaceutical sectors are not just looking for "code writers," but AI orchestration specialists, AI security experts, and data analysts who will validate the results of autonomous systems.
According to an IEEE survey, up to 96% of technologists expect rapid adoption of Agentic AI. However, as Gartner warns, the costs of operating AI coding itself could exceed the salaries of software developers by 2028. This means that efficiency is necessary, but managing these systems will be costly.
What Does This Mean for the Czech Market and European Companies?
For Czech software houses and technology startups, this situation comes with a dual face:
- Competitive Pressure: Global companies using Agentic AI will be able to deliver software faster and cheaper. Czech companies will not find it profitable to specialize in "manual" writing of repetitive code, but must adopt these tools to maintain their margins.
- Regulation and Security (EU AI Act): Within the European Union, adherence to rules for high-risk systems will be crucial. Autonomous coding in critical infrastructure (e.g., healthcare or banking) is subject to strict rules for transparency and oversight. A Czech company must ensure that the AI agent does not generate code that violates GDPR or EU security standards.
Prices and Availability of Tools
If you want to start experimenting, here is an overview of current options:
- GitHub Copilot: Standard for most developers. The price is around 10 USD/month for individuals (there is also a free tier for students). It is fully available and works within standard IDEs.
- Devin (Cognition AI): Focused on the enterprise segment. The price is not public, requiring direct contact with sales, but it is intended for large teams solving massive migrations.
- Open-source alternatives (e.g., OpenDevin): For those who want full control over data (which is crucial in the Czech Republic due to GDPR), there are open-source projects that can be run on your own server.
Availability in Czech: Most of these tools communicate primarily in English. For a Czech developer, this is not an obstacle for the code itself, but it can be a barrier for complex documentation or when working with local legal regulations. We recommend using models with a high ability to understand context (like Claude 3.5 Sonnet), which handle Czech very well.
Will Agentic AI completely replace human programmers?
No, but it will radically change their role. The programmer becomes an architect and controller (reviewer). Instead of writing every line, they will define goals, check security aspects, and solve complex logical problems that AI cannot yet fully comprehend.
Is it safe to let AI generate code for banking systems?
Only under strict human supervision (Human-in-the-loop). Companies must implement processes where every line of code created by an agent undergoes human validation and automated security tests to prevent vulnerabilities.
What are the main costs of implementing Agentic AI in a company?
In addition to the subscriptions themselves (SaaS), companies must account for infrastructure costs (computing power for running agents), employee training costs, and especially the costs of security auditing the generated code.
