Right now, in coding, we operate within two distinct ranges of execution methods. On one end, we have traditional AI assistance - tools like ChatGPT integrated into sidebars that provide suggestions and guidance while developers maintain complete control. On the other end, we have fully autonomous coding agents that can independently write, test, and iterate on code with minimal human oversight.
We're entering a new phase where the industry is shifting toward a more autonomous modality. The old paradigm, where individual developers handled most tasks manually, is evolving into one where programmers increasingly orchestrate AI agents that act as collaborative contributors. This shift represents a fundamental change in how we approach software development.
Here's how the workflow for creating a new feature used to look for many coders:
- Write code in VSCode.
- Test the code.
- Commit and push to Git (with checks like linting, pre-commit hooks, and automated tests).
- Changes were reviewed (by the team) and possibly modified, then merged into the main branch.
- Continuous integration to production.
Now, things are different.
- Coding can be done using tools that can be grouped into two categories:
AI-enhanced editors - Cursor, Zed, Windsurf, VSCode with Copilot, Trae
Agents - Claude Code, Codex, Aider, Goose, Replit, Devin
Since AI agents can now create reasonably reliable code, you can assign them to work on different features at the same time. For example, I can manage multiple features using Conductor while also working on others with the agents in Cursor.
- Testing is simpler. AI models excel at writing tests, especially unit tests, because they play to their strength in recognizing patterns.
- See point 2.
- AI tools like Coderabbit and Greptile can now review your code changes. I prefer using a "council" of AI reviewers for my personal projects.
The bottleneck now is attention. Output is limited by the number of agents or threads that can be monitored and reviewed. This also includes managing the various components of this new ecosystem, such as model configuration presets (MCPs), skills, hooks, and prompts. It's a bit chaotic now, but that's where the excitement is. The AI models and tools supporting them will continue to improve, solidifying this new programming paradigm.
If you're not using AI, you're falling behind. But if you rely on it too much, you'll also be left behind.
(Inspired by https://x.com/karpathy/status/2004607146781278521?s=20)