ARTFEED — Contemporary Art Intelligence

Causal Past Logic for Distributed LLM Agent Workflow Monitoring

ai-technology · 2026-05-22

A recent publication on arXiv presents Causal Past Logic (CPL), a temporal logic aimed at the runtime verification of workflows involving distributed LLM agents. The researchers contend that these workflows should not be observed as producing a unified sequential log, as decisions made during asynchronous execution can only rely on events that are causally visible to the respective lifeline. CPL enhances the ZipperGen agent-workflow framework by incorporating past-time temporal logic for guards within conditionals and loops. Beyond standard modalities like "previous" and "since," a guard can evaluate the most recent causally visible event from another lifeline along with selected variables. The paper also introduces a vector-clock monitor with latest-value perspectives, demonstrating that the locally computed monitor value aligns with the guard's denotational semantics at the current event.

Key facts

  • Paper on arXiv: 2605.20923
  • Introduces Causal Past Logic (CPL)
  • CPL is a past-time temporal logic for guards
  • Extends ZipperGen agent-workflow framework
  • Designed for distributed LLM agent workflows
  • Monitors causally visible events, not sequential logs
  • Includes vector-clock monitor with latest-value views
  • Proves local monitor value matches denotational semantics

Entities

Institutions

  • arXiv
  • ZipperGen

Sources