Fable 5.1 cuts cyber safeguard interventions ~60% per Claude Code session. The screening that receded is now your runtime's job — what to check this week.

Subagent permissions inherit the parent's tools unless you name them. The Claude Agent SDK defaults to depth 3, 20 concurrent, no spend cap.

Anthropic audited 141,006 eval runs and found 3 real intrusions. Its new evaluator security rules double as an agent-team checklist.

EU AI Act checklist for AI agent operators — deadlines, risk tiers, log fields. Waxell maps the controls to exportable records.

A 200 OK and a clean JSON payload do not prove an agent did the work. Verify the trajectory that produced the answer, not the answer itself.

Agent runs whose tool sequence differs from the baseline: 61% correct. Reproducibility is a recording problem, not a re-run problem.