Two months ago, I was talking about how we would “direct” armies of AI agents. Today, we are learning the hard way that an infinite loop not only crashes the processor, but also empties your bank account.

When I last posted on LinkedIn about multi-agent systems, I was in the “honeymoon” phase. Everything seemed like a natural transition from microservices to autonomous agents. But the reality of production over the last 8 weeks has hit us with some hard lessons about what Agentic Engineering means.

1. FinOps has become the new critical hard skill. In classic programming, a bug in a while(true) loop would crash your application. In the world of agents, two agents entering a loop of pleasantries (“Thank you!” / “You’re welcome!” / “No, thank you!”) can generate thousands of dollars in API credits overnight. Engineering is no longer just about logic, but about financial circuit breakers and optimizing the cost per task.
2. Hard skills are not disappearing, they are evolving. It has been said that AI will abstract complexity. False. Complexity has only shifted. Orchestrating non-deterministic agents (which don’t always give the same answer) requires a deep understanding of distributed systems. State management, race conditions between agents, and security (not allowing them to accidentally delete the database) are problems that no LLM can solve on its own. We need architects, not just “prompt engineers.”
3. What can we expect in H1 2026? We are moving from the “Exuberance” phase to the “Responsibility” phase. The next six months will no longer be about what agents can do (wow demos), but about how much it costs us to do so. Companies will start auditing: does the agent bring real ROI or is it just an expensive experiment?

We are not losing our jobs, but our job descriptions are being rewritten in real time. We are moving from writing code that executes to building systems that think – and that is much harder than it sounds.

– Pavel Covali (.Net Tech Lead)