Built a demo where an agent can provision 2 GPUs, then gets hard-blocked on the 3rd call

About this article

Policy: - budget = 1000 - each `provision_gpu(a100)` call = 500 Result: - call 1 -> ALLOW - call 2 -> ALLOW - call 3 -> DENY (`BUDGET_EXCEEDED`) Key point: the 3rd tool call is denied before execution. The tool never runs. Also emits: - authorization artifacts - hash-chained audit events - verification envelope - strict offline verification: `verifyEnvelope() => ok` Feels like this is the missing layer for side-effecting agents: proposal -> authorization -> execution rather ...