A runaway AI agent can create enormous enterprise consequences even when each individual operation is technically permitted.
The control problem is not simply whether Agent A may call API B. Enterprises also need to constrain how many times, for how long, at what cumulative cost, under what termination condition and across how many subordinate agents.
Authority needs budgets
Authority Budgets should be hierarchical: enterprise to mandate to swarm to agent to task to effect.
A subordinate agent can consume a delegated portion of a parent budget but cannot silently create additional budget.
Useful dimensions include financial cost, token and API usage, wall-clock duration, action count, compute, data volume, resource mutations, downstream-agent creation and cumulative consequential effects.
Swarms make the problem harder
This becomes especially important at swarm scale. Ten agents cannot each inherit the entire parent budget unless the enterprise explicitly authorized that aggregate effect.
Delegation can divide authority. It cannot silently multiply it.
Sources
NEXT STEP