Containers, microVMs or gVisor for AI agents?
How containers, gVisor and Firecracker-style microVMs isolate an AI agent, what each boundary costs, and which to pick when the agent runs untrusted code.
Research notes
Research notes
All notes, page 2 of 2. The guides by topic are on page one.
Latest research notes
All notes, newest first
Every note answers one question, cites its sources, and links to the guide it belongs to.
Logs vs evidence: what an auditor can trust
An AI agent's own log is a claim, not evidence. The four properties a record of agent actions needs before an auditor, regulator or insurer can rely on it.
EU AI Act Article 12: logging for AI agents
Article 12 of the EU AI Act makes high-risk AI systems allow automatic logging of events. What that means when the system is an agent, and when it applies.
AI Omnibus: new dates for high-risk obligations
The AI Omnibus moved the EU AI Act's high-risk duties to 2 December 2027 for standalone systems and 2 August 2028 for embedded ones. What that changes.
Impossible tasks and what exploit rates reveal
Why evaluations include tasks that cannot be completed, how ImpossibleBench measures the rate at which agents pass them anyway, and what the number means.
How to read an AI agent leaderboard
How to read an AI agent leaderboard: task families, two rates instead of one score, confidence intervals in plain language, versions, dates and who ran it.
Benchmark contamination: why public scores decay
Why AI agent benchmark scores rise without the skill rising: how tasks leak into training data, what ARC Prize found, and what a private set changes.