Writing

Notes on AI systems, agent architecture, quantitative modelling, and what I learn shipping them. Posts appear here first, then go out on Substack.

Your Safeguards Govern Tools, Not Your Agent's Behavior

April 19, 2026

Your safety controls see tool names and your agent sees goals

AI AgentsAI SafetySecurityDragonfly

I got tired of giving daily context to my AI assistant. So I stopped.

April 16, 2026

Giving Claude a sense it's never had.

AI AgentsMCPClaudeGeolocate Me

Your AI Agent's Safety Controls Don't Fail Under Pressure

April 15, 2026

440 experiments, two models, and the variable that actually matters isn't what we expected.

AI AgentsAI SafetyResearchDragonfly

Controlling an AI Pentesting Agent: What Actually Works and What Doesn't

April 4, 2026

The difference between governing what your agent does and governing how it does it.

AI AgentsAI SafetyPenetration TestingDragonfly