RelLedger audits shared facts in AI agent organizations
RelLedger: Relational Information Governance for LLM Agent Organizations
Multiagent SystemsComputation and LanguageComputers and Society
Summary
In organizations run by large language model (LLM) agents, shared records guide later decisions, but it can be unclear who influenced what was recorded and how it affects future steps. The authors developed RelLedger, a tool to trace these connections by checking different interaction histories and event roles to find the key factors controlling decisions. By testing multiple AI models, they found distinct patterns of information sharing and influence depending on the history context. Their system helps understand how AI agents communicate and make decisions based on shared facts.
What this means in practice
- •For ai system developers: Evaluate and audit how AI agents share and act on information in multi-agent environments to improve trust and transparency.
- •For workflow automation engineers: Use executable workflows to trace information flow and outcomes within automated company processes involving multiple AI tools.
Authors
Junxi Huang, Wenshuo Chen, Nuoxuan Li, Kailong Wang, Meitong Shen, Yuehan Kou, Leihao Zhao, Yutao Yue, Xiucheng Wang, Nan Cheng
Abstract
In LLM agent organizations, shared records determine which facts are available for later decisions. A sender consults interaction histories before deciding whether to report an event. A delivery gap across histories does not show whose relationship drove the choice or whether the resulting record matters downstream. We introduce RelLedger to audit these links. It holds the event, evidence, risk, and message fixed while crossing the sender's histories with the event target and an unrelated actor. Role reassignment and history rewrites locate the controlling cue. Message and cost interventions show when the effect appears. An attribution ledger maps each claim to its required response cells and separates missing interventions from sampling uncertainty. Recipient replay follows delivered information into downstream decisions. Executable workflows for a simulated one-person company extend the audit to retrieved histories, tool calls, and shared ledger updates. Across GPT, Gemini, Claude, DeepSeek, and Qwen, the audit reveals model-specific control paths and smaller relation gaps under explicit loss rules. Claude sends the same factual notice 62.22 percentage points more often with conflict histories than with friendly histories. In executable workflows, strained histories reduce recipient review rates by 45.83 points relative to warm histories for Claude.