A Probabilistic Model for Detecting and Preventing Agent Drift

aming5571 pts0 comments

The Charting Loop: A Four-Layer Probabilistic Theory of Uncharted-to-Charted Work in Agent Systems | Zenodo

Skip to main

You are using an outdated browser. Please upgrade your browser to improve your experience.

Published August 8, 2026

| Version v1

Preprint

Open

The Charting Loop: A Four-Layer Probabilistic Theory of Uncharted-to-Charted Work in Agent Systems

Authors/Creators

Zhang, Ying1

Show affiliations

1.

Independent Researcher

Description

Long-horizon agent work is usually framed as task execution. We propose charting as a more general unit of analysis. A problem is uncharted for an executor when progress requires changing its representation, objective, or admissible action structure; it becomes charted when a stable position vocabulary, direction criterion, transition policy, and evaluation procedure exist for it — held as skill, routine, and institution in human executors, and, in governed agent runtimes, compilable into machine-checkable constraints and executable transitions. Work, on this view, is the conversion of uncharted problems into charted corridors — corridors that are then walked repeatedly with novel content. At recursion layer ℓ we model a navigation-valid step as the conjunction of a valid position, a valid direction, and a valid entrance, giving Pr(N_ℓ) = Pr(P_ℓ) Pr(D_ℓ | P_ℓ) Pr(E_ℓ | P_ℓ, D_ℓ) with entrance further decomposed into availability and selection and execution fidelity explicitly factored out of the theory's scope. We organise the account as a four-layer compilation stack — phenomenon, theory, system, exogenous authority — whose system layer compiles the factors into runtime constraints: verified state, frozen acceptance data, pushed single entrances, warranties. The dynamics form a charting loop: an exogenous authority — in our instantiation, a human operator — provisions intent, domain knowledge, rules, and trusted evidence sources; executors build corridors, walk them, repair them, and certify them by fresh end-to-end traversal; new situations re-enter the loop, charted either by the authority through reframing and redirecting or, where the compiled theory suffices, by the executor itself subject to ratification. The loop is recursive in two senses: its outputs re-enter it as objects, and every revolution — at whatever meta-level — is itself work obeying the factorization. We illustrate the framework with traceable incidents from a production governed multi-agent runtime, treating them as motivating cases rather than empirical validation, and we derive falsifiable predictions for long-horizon drift, repair, agent self-extension, and machine-authored governance.

Files

charting-loop-zenodo.pdf

Files<br>(220.4 kB)

Name<br>Size

Download all

charting-loop-arxiv.tar.gz

md5:90c0a1df5104c01ecf620b02bfeafeca

31.5 kB

Download

charting-loop-zenodo.pdf

md5:c24af24faab0e6500a16057797323f4c

189.0 kB

Preview

Download

228

Views

13

Downloads

Show more details

All versions<br>This version

Views

Total views

228

228

Downloads

Total downloads

13

13

Data volume

Total data volume

2.0 MB<br>2.0 MB

More info on how stats are collected....

Versions

External resources

Indexed in

OpenAIRE

Communities

Keywords and subjects

Keywords

multi-agent systems

agent governance

normative multi-agent systems

long-horizon agents

AI infrastructure

norm synthesis

runtime verification

organizational routines

Details

DOI

DOI Badge

DOI

10.5281/zenodo.21844624

Markdown

[![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.21844624.svg)](https://doi.org/10.5281/zenodo.21844624)

reStructuredText

.. image:: https://zenodo.org/badge/DOI/10.5281/zenodo.21844624.svg<br>:target: https://doi.org/10.5281/zenodo.21844624

HTML

Image URL

https://zenodo.org/badge/DOI/10.5281/zenodo.21844624.svg

Target URL

https://doi.org/10.5281/zenodo.21844624

Resource type<br>Preprint

Publisher<br>Zenodo

Languages

English

Rights

License

Creative Commons Attribution 4.0 International

The Creative Commons Attribution license allows re-distribution and re-use of a licensed work on the condition that the creator is appropriately credited.

Read more

Copyright

Copyright (C) 2026 Ying Zhang. Licensed under CC BY 4.0.

Citation

Export

Technical metadata

Created

August 7, 2026

Modified

August 7, 2026

Jump up

This site uses cookies. Find out more on how we use cookies

Accept all cookies<br>Accept only essential cookies

zenodo agent loop charting work https

Related Articles