Blog
Notes from the boundary.
What we learn building an authorization layer that sits outside the agents it governs.
- 2026-10-06
What Actually Happens When an AI Coding Agent Writes Through a Symlink?
Six traced runs of Claude Code, Gemini CLI and Codex CLI appending one line to a file that was actually a symlink: what the system-call trace shows about intent, authorization, resource resolution and execution.
- 2026-09-25
What Four Coding Agents' Hook Systems Actually Do
Where an agent's action can be intercepted in Cursor 3.21, Gemini CLI 0.60, Codex CLI 0.155 and Claude Code 2.1, what a hook can answer, and how each runtime fails. Version-specific observations, not compatibility claims, from wiring one authorization layer into all four.