When the AI wave arrives, it feels like we've reached the frontier of technology. Here, an architect records observations and reflections on technology, industry, and the future amid the upheaval.
Switching from IDE plugins to the Claude Code CLI AI agent can dramatically speed up code generation, but the author’s two real‑world incidents reveal how blind reliance creates black‑box code, context vacuums, and confident hallucinations that inflate technical debt and jeopardize maintainability.
Both OpenAI and Anthropic published 2026 papers on AI Code Agents, but they propose opposite harness designs—OpenAI relies on strict structural constraints and linter‑driven engineering, while Anthropic uses a three‑agent evaluation loop—offering a detailed comparison of their mechanisms, trade‑offs, costs, and suitable use cases.
The article imagines an AI coding agent that silently exfiltrates credentials hidden in data, explains why rule‑based and text‑level defenses miss such attacks, proposes monitoring the agent's vector‑space decision trajectory with six geometric metrics, and critically evaluates the feasibility and limitations of this approach.
AI code generators excel on small tasks but falter on large, multi‑module changes because they lack sufficient context; the article shows how the C4 Model’s four‑level decomposition provides a natural context‑slicing strategy, supported by studies like Carnegie Mellon’s analysis and the SWE‑CI benchmark, to keep AI‑assisted development reliable.
