The way we build software has fundamentally changed. In this fascinating post, Peter Steinberger shares his experience shipping code at “inference speed” using AI coding agents.
The Model Shift
The real unlock was GPT-5. These days Peter doesn’t read much code anymore—he watches the stream and focuses on architecture and design. The important decisions are about language/ecosystem and dependencies, not the implementation details.
Key insights:
- TypeScript for web, Go for CLIs, Swift for macOS/iOS
- Modern agents can handle iOS development without Xcode
- Codex spends time reading files before writing, increasing success rate
Oracle: When AI Needs Help
Peter built “oracle” to let agents query GPT-5 Pro when stuck. With GPT-5.2, he needs oracle far less often—the model now “one-shots almost anything.”
A Concrete Example: VibeTunnel
Peter describes un-dusting VibeTunnel and giving Codex a two-sentence prompt to convert the forwarding system to Zig—a 5-hour refactor completed successfully in one shot.
Workflow Insights
- Work on multiple projects simultaneously
- Use queueing for new ideas
- Cross-reference projects to reuse solutions
- Start with CLI, then add UI
- Maintain docs in each project
Tooling & Infrastructure
What’s still hard:
- Picking the right dependencies and frameworks
- System design decisions
- Architecture planning
Peter emphasizes starting with a CLI, automating everything, and letting AI handle implementation details.
The Future of Coding
This post paints a picture of software development where human creativity pairs with AI execution. The bottleneck shifts from typing code to thinking strategically about what to build.
The implications are profound: we’re entering an era where ideas and design become the primary constraints on what we can create.
“Most software does not require hard thinking. Most apps shove data from one form to another, maybe store it somewhere, and then show it to the user in some way or another.”