AI. What’s next?

Up until around December, using AI in development basically meant: prompt -> copy some code -> paste -> tweak -> repeat. Now this already feels outdated and clumsy. Because we now have tools that don’t just suggest, they actually do the work:

1. Claude Code – an agent that walks through your files, understands context, edits code, runs tests, and can drive a task to definition of done.
2. Codex – it now has dedicated development modes/models. For example, gpt-5.3-codex is not just 'answer with code' but 'let’s solve a real project task with iterations and tools'. It can keep context over long iterations, and context window overflow is no longer the end of the dialogue, because the model manages context itself and can compress it with almost no quality loss.
3. And probably the freshest but no less interesting one is OpenClaw. It is essentially an orchestrator of many agents: one researches, another writes code, a third reviews, a fourth fixes tests, a fifth prepares the release. All of this can be automated almost down to you stating the goal in chat and, after a few iterations, getting a working product.

Now the unpleasant part:
For a long time people in IT lived in a paradigm: 'I will learn the fundamentals, I will learn more tools, and my job and prospects will be fine.' That is the first thing that has changed. Knowing applied implementation, or at least understanding it, is still important, but it is no longer the decisive factor for success.

So a small piece of advice: stop grinding LeetCode problems. Right now, other things matter much more:

• abstractions, interfaces, and system boundaries
• architecture
• observability, operations, and security
• and of course the ability to formulate tasks so that they can be executed not only by a human but also by an agent

I also have something for companies:
If by 2026 your AI 'adoption' is still at the level of 'ask ChatGPT' plus setting up Copilot autocomplete, and your SDLC is not at least partially agentized, I have bad news for you. Starting from 2027, you will be losing not only to faster competitors but also to individual enthusiasts who have wired up a mesh of autonomous agents that do more in one week than your department does in a month, because 90% of your time people are just writing code and endlessly delivering features.

The world has changed. It is time to act.