A seasoned Principal Engineer recently completed a full iOS app using only AI agents, working just 1-2 hours daily for two months, but discovered that speed and efficiency came at an unexpected cost: the fundamental satisfaction of engineering itself. The experiment with Claude Code, OpenAI's Codex, and other AI tools revealed a troubling paradox in the future of software development, where productivity gains may come at the expense of technical growth and professional fulfillment.

What Happens When Engineers Stop Coding?

The engineer, who had 20+ years of professional experience and considered himself highly skilled across multiple programming languages and paradigms, set out to test whether AI agents could truly replace human developers. His goal was straightforward: create something useful that he had zero experience with, without ever opening a code editor or understanding what the AI agents were actually writing.

Using Claude Code with Opus 4.6 and Sonnet 4.6 models, along with OpenAI's Codex and Copilot CLI, he successfully built "Papers," a document management iOS app now available on the App Store. The project was completed entirely through prompting and orchestrating AI agents, with the engineer acting as a director rather than a builder. However, the experience revealed something unexpected: the absence of hands-on coding fundamentally changed how he thought about engineering.

"I got better as a leader and an architect, but I'm not sure I got better as an engineer in any way. The main reason is that even for ideating I used AI, so I didn't spend much brain-time actually thinking," the engineer reflected on his experience.

Principal Engineer, Aliz Tech

How to Effectively Manage Multiple AI Coding Agents

The engineer developed a systematic approach to orchestrating AI agents that proved essential for success. Rather than relying on a single tool, he employed a cross-checking methodology that helped catch errors and inconsistencies.

• Parallel Task Distribution: Working with a maximum of four agents simultaneously on separate Git branches, ensuring tasks didn't overlap and each agent had clearly defined responsibilities without confusion.
• Cross-Verification Process: Creating the first version with one tool, then cross-checking implementation with the other two tools to identify discrepancies and potential issues before testing.
• Iterative Refinement Cycles: Conducting multiple rounds of audits and security checks, reverting problematic iterations rather than layering fixes on top of existing flawed code.
• Architectural Oversight: Using best practices from 20+ years of experience to force agents to check their own work, audit security, and challenge instructions when they seemed incorrect.

Which AI Coding Tool Performed Best?

Claude Code emerged as the most capable tool in the engineer's assessment, though each platform had distinct strengths and weaknesses. The comparison revealed important differences in reliability and output quality that matter for production applications.

Claude Code and OpenAI's Codex performed at similar capability levels, but GPT models generated significantly more "slop," a term the engineer used to describe unnecessary, redundant, or poorly structured code that accumulated through iterations. However, Claude Code had a critical vulnerability: when Anthropic experienced model regression problems, Claude's performance degraded dramatically, becoming unreliable for complex tasks.

The engineer's experience highlights a fundamental tension in agentic development: while AI agents can accelerate project completion, they introduce new dependencies and failure modes that traditional software development doesn't face. Provider outages, model regressions, and the non-deterministic nature of large language models (LLMs) created situations where the engineer had zero ability to continue without external AI services.

Can You Use Claude Code for Free?

While Claude Code itself is free and open source, the actual cost comes from API calls to the underlying language model. However, developers can now access Claude Code's full functionality without paying Anthropic's subscription fees by using local models through Ollama, an open-source tool that runs AI models directly on personal hardware.

Claude Code works by acting as an intermediary layer that manages file editing, terminal commands, and project context. The actual thinking and code generation happens in the underlying model, whether that's Claude Sonnet, Claude Opus, or any other language model. By pointing Claude Code to a local Ollama instance instead of Anthropic's servers, developers can use the same interface and capabilities completely free.

The trade-off is model quality. Open-weight models like Qwen 3.6 and Gemma 4 don't match Claude Opus's capabilities, but for everyday coding tasks, the gap has become surprisingly small. Qwen 3.6, built specifically for agentic coding, comes in 27 billion and 35 billion parameter variants and requires 17-24 gigabytes of memory. Gemma 4's 26 billion parameter mixture-of-experts model uses only 4 billion active parameters, making it more efficient while still delivering solid coding benchmarks.

For developers with modern Apple Silicon Macs featuring 32 gigabytes of unified memory, or those with high-end GPUs, running these models locally is practical. Even those with 16 gigabytes of memory can use Gemma 4's smaller variant, which runs in approximately 5 gigabytes in 4-bit mode.

The Hidden Cost of Agentic Development

The engineer's most important finding wasn't about technical capability, but about professional development and job satisfaction. While agentic development is undeniably powerful for shipping products quickly, it fundamentally changes what it means to be an engineer.

The experience revealed several concerning patterns. First, the engineer never achieved a "flow state" even when working with just two agents. The constant context-switching between prompting, reviewing, and orchestrating created heavy stress rather than the satisfaction typically associated with creative technical work. Second, the accumulation of "slop" became increasingly dangerous as iterations continued without cleanup, requiring the engineer to revert entire sections and restart rather than iterate forward.

Most significantly, the engineer noted that he missed "the joy of coding" itself. The entire two-month period focused exclusively on efficiency, speed, and productivity metrics. The creative problem-solving and experimental thinking that typically characterizes engineering work was largely absent because the AI agents handled ideation, implementation, and iteration.

"I felt I was getting closer and closer to the movie 'Idiocracy.' The more you iterate without cleaning up between iterations, the more slop creeps in," the engineer observed about the accumulation of technical debt in agentic workflows.

Principal Engineer, Aliz Tech

The engineer's conclusion suggests that while agentic development will become standard practice, the industry's approach to it may be misguided. Rather than viewing AI as a tool to eliminate the need for human engineering judgment, he advocates for using AI to augment engineering while preserving the technical growth and professional satisfaction that makes engineering meaningful.

As AI coding tools become more capable and accessible, both through paid services like Claude Code and free alternatives using local models, the real question isn't whether developers can build faster. It's whether the speed gains are worth the cost to professional development and the engineering craft itself.