Fault-Tolerant Budget Conservation for Distributed Multi-Agent Delegation

Summary

The paper studies how resource limits can remain an authorization boundary when AI agents delegate work across concurrent workers that may fail. It formalizes budget conservation using quantized resource vectors represented as exclusive escrow credits moving through a delegation DAG. Before dispatch, a branch converts credit into an operation reservation tied to its lineage, epoch, normalized effect, maximum charge, receiver, and idempotency key, then stores a signed dispatch permit in quarantine for gateway verification. Effects whose outcome is uncertain remain charged until authenticated settlement, an authoritative proof that no effect occurred, or permanent retirement. Under stated mediation, durability, authentication, normalization, and gateway assumptions, the authors prove ownership partition, ledger and effect conservation, non-amplification by descendants, at-most-once settlement, safety for late completions, and confinement during network partitions. An indistinguishability result shows that availability during a partition requires exclusive preallocation. Bounded TLA+ checks, an independent JavaScript explorer, and crash-injected two-process SQLite experiments tested the declared scope, including timeout refunds, retries, duplicates, joins, partitions, and late completions. The evaluated schedules preserved the issued budget bound.