AI sub-agents inherit security limits from their parent

A policy algebra derives each AI sub-agent's security limits from its parent's, per the arXiv paper.

That closes a real gap: guardrails get hand-copied into the first few agents, but the eleventh one someone spins up later doesn't get updated. The composed system intercepts 94.8% of policy-violating events, keeps 86.9% of task completion, and logs 98.6% audit completeness, per the arXiv paper.

Security profiles and runtime obligations compose through joins, intersections, budget narrowing, and approval inheritance. A child agent's limits get calculated from its parent's, instead of typed in by hand for every new agent. The composition rules carry formal correctness conditions and executable semantics, per the arXiv paper.

Budget narrowing is the part worth sitting with. It's the same idea as putting a spend cap on an agent, made composable. A sub-agent's budget derives from what its parent was allowed, rather than set separately and forgotten.

The completion cost is the real number

The interception rate will get the attention, but 86.9% task completion decides whether anyone adopts this. A security layer that blocks 94.8% of violations while costing 13% of completed tasks is a hard sell to a team on deadline. The arXiv paper's formal guarantees don't say whether that tradeoff holds outside its own test conditions. Watch for whether this ships as a library instead of staying a spec. That's the difference between fixing guardrail drift for real and one more paper nobody retrofits onto agents already running.