Budgeted Audit-Point Placement

An auditor agent is scheduled across a running multi-agent workflow modelled as a dynamic dependency graph: under a fixed audit-call budget, audit points are placed greedily on influential and long-unaudited regions to minimise cumulative unchecked exposure, and each trusted audit point becomes a recovery anchor so failures roll back locally instead of to the start.

Properties

family: Evaluation & Ranking
maturity: frontier
diagram: judge
whyMultiAgent: The quantity being minimised, cumulative unchecked exposure, is defined over inter-agent dependency edges through which one agent's hallucinated or malicious output propagates into another agent's work; a single agent has no dependency graph to audit and no propagation to bound. Distinct from the already-tracked llm-as-judge pattern, which scores a finished output against a rubric: here the design variable is where and when in a live multi-agent execution the auditor attends, under an explicit budget, not how any one output is scored. Also distinct from reasoning-tree-audit, which adjudicates divergences between completed independent reasoning traces post-hoc: audit-point placement happens during execution over a dependency graph of collaborating specialists, and its trusted checkpoints exist to localise rollback, a runtime concern that post-hoc trace adjudication does not have.
mast: [Cat 2][Cat 3]
useCases: [Guarding agent pipelines against propagated hallucinations at bounded token cost][Localised rollback and recovery in long multi-agent trajectories][Runtime detection of a compromised or malicious agent in a collaborating team]
source: BRA-Audit · arXiv:2608.14668

Provenance

Source: https://arxiv.org/abs/2608.14668

Budgeted Audit-Point Placement

Properties

family

Evaluation & Ranking

maturity

frontier

diagram

judge

whyMultiAgent

The quantity being minimised, cumulative unchecked exposure, is defined over inter-agent dependency edges through which one agent's hallucinated or malicious output propagates into another agent's work; a single agent has no dependency graph to audit and no propagation to bound. Distinct from the already-tracked llm-as-judge pattern, which scores a finished output against a rubric: here the design variable is where and when in a live multi-agent execution the auditor attends, under an explicit budget, not how any one output is scored. Also distinct from reasoning-tree-audit, which adjudicates divergences between completed independent reasoning traces post-hoc: audit-point placement happens during execution over a dependency graph of collaborating specialists, and its trusted checkpoints exist to localise rollback, a runtime concern that post-hoc trace adjudication does not have.

mast

[Cat 2][Cat 3]

useCases

[Guarding agent pipelines against propagated hallucinations at bounded token cost][Localised rollback and recovery in long multi-agent trajectories][Runtime detection of a compromised or malicious agent in a collaborating team]

source

BRA-Audit · arXiv:2608.14668