Reading Stability from Traces: Certifying Multi-Agent LLM Loops with Unknown Delay from Short, Contaminated Execution Logs
Third paper in the series "Control Theory for AI Systems". Language-model agents that correct a shared state form a delayed feedback loop; prior work gave its exact stability boundary and showed that prompting does not control the agents' gain. This paper shows how to certify the stability of a running loop from a short execution log, with the delay unknown. The certifier combines a scale-invariant weighted-median estimator, a bounded delay-selection loss and a bootstrap that re-selects the delay, with five proven guarantees. It is evaluated on 2,880 synthetic traces, a model-mismatch study, 54 recorded runs of gpt-4o-mini agents and 4,050 production latencies, against least squares, Huber regression, ARX and delay-blind regression. The archive contains the LaTeX source, all code, data and raw outputs, the data-collection scripts, and a test suite (177 tests) that checks each proposition numerically, regenerates every number and table of the paper, and links every number typed in the text to the code or data. A mutation test plants 15 realistic bugs and verifies that the tests catch all of them.
Authors
- Mohamed EID (ORCID: https://orcid.org/0009-0008-8404-978X)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-17
- DOI
- https://doi.org/10.5281/zenodo.24322429
- Primary Topic
- Stability and Control of Uncertain Systems
- Type
- preprint