6 hours ago
- Coding agents rapidly gained popularity among software developers after Claude Code's surge in late 2025, with figures like Andrej Karpathy noting a shift away from manual coding since December 2025.
- Grady Booch views this as the third golden age of software engineering, where AI agents abstract code itself, similar to how compilers abstracted the machine.
- The core problem with 'vibe coding' is the loss of understanding, awareness of dead code, and the ability to reason about or explain architectures, leading to unreliable outputs.
- Current solutions like markdown specs, skills, Spec Kit, OpenSpec, Kiro, and test-driven development fail to address trust, offering only prompt engineering without deterministic enforcement.
- Promising approaches like CodeSpeak and Scryer focus on trust by compiling specs into code or creating a shared model between developer and agent for transparency and drift detection.
- The future likely involves moving beyond manual coding for routine tasks, with agents guided by structured, auditable intent rather than vague prompts.