From Prompts to Agents: How a Doubletapp Engineer Reformatted the Approach to AI

Andrey Zharov, an iOS developer at Doubletapp, shared an experience of evolution that reflects the path of the entire industry. First came simple magical prompts like "make no mistakes" and "write if you need more context," which almost magically pulled working code from ChatGPT. Then came a phase of agent euphoria, when it seemed like you could automate absolutely everything. But ultimately, he came to a conclusion: a completely different approach was needed. Not so much the art of writing perfect prompts, but the engineering of building reliable AI infrastructure.

From Magic to Systematics

The first period was the era of vibe coding. Open ChatGPT, write a task, get chunks of code, assemble it into an app. If you're lucky, it works. Enable Deep Thinking—the result is more stable. It seemed like we found a magic wand. But reality was harsher. Results turned out to be unstable. Prompts grow in size, but reliability falls. The same task is solved differently on different days. A completely different approach was needed—not asking the model for code, but systematizing how we interact with it.

The industry evolved. They moved from simple chat to local agents, then realized the importance of orchestration, not just chains of calls. And at some point, everyone collectively understood: we should not write code in ChatGPT. Instead, we should write prompts that write correct prompts. And build infrastructure that works reliably with these prompts.

Context Engineering as Foundation

The key point is context engineering. It's not just "give more context and the model will understand." It's a systematic construction of the architecture of what information and in what order the model sees, what tools are available to it, and how they're connected to each other.

Important elements:

• Proper context preparation is half the success of an agent
• The toolset must be predictable and safe for the model
• Error handling and rollback should work like in regular code
• Testing and monitoring are not an option, but a requirement
• Versioning of prompts and configs, like in Git

This is no longer the art of writing a super-prompt. This is engineering.

Orchestration Instead of Magic

Then comes orchestration. Not one big agent that does everything at once. But a system where each component is responsible for one part. One parses context, another works with logic, a third validates the result. It's similar to regular application architecture, but applied to AI systems.

And here an interesting paradox appears: you start worrying about prompt injection not as an exotic vulnerability, but as part of your security, like you once did with SQL injection. You need guardrails, input validation, clean interfaces between components—all like in regular development.

Then comes meta-prompting—when the model itself helps you generate and optimize prompts. But this is no longer magic, but a tool.

What This Means

Developers are moving to a new level: from "ask ChatGPT for code" to "build a system that knows how to work with AI reliably." AI development becomes more predictable, scalable, and technically sound. Like regular development, but with new challenges.