Published May 5, 2026 | Version v2

OAgents: A Pre-Standardization Draft Profile for Operational AI Agent Trustworthiness

Description

Large language models are increasingly deployed as operational agents that manage infrastructure, generate enterprise content, and execute multi-step workflows. Capability has advanced faster than trust mechanisms. This document proposes OAgents, a community-authored pre-standardization draft proposing a behavioral control and evidence profile for operational AI agent trustworthiness.
OAgents defines a behavioral envelope: a structured set of pre-execution gates, post-execution verification controls, and operational-discipline mechanisms. It specifies 26 controls across 7 categories, 3 conformance levels, and observable evidence criteria. The profile is model-agnostic. Behavioral guarantees are properties of the envelope, not of the underlying model.
The draft is structured as an AI RMF Implementation Profile, mapping its control taxonomy to NIST AI RMF 1.0 functions, categories, and subcategories. A reference implementation demonstrates architectural feasibility for all 26 controls in a development-laboratory environment with live operational use. No generalized empirical benchmarks are claimed. Broader validation across diverse deployment contexts remains community work.
This document is offered for technical review, comparison, and contribution to future AI agent standards development. It is intended to inform ongoing work at NIST, NCCoE, IEEE, OASIS, and related venues. It is not a finalized or adopted standard.
A sanitized reference implementation is available at: https://github.com/ologos-corp/OAgents-standard

Files

oagents-nist-standard-v16.0.pdf

Files (675.2 kB)

Name Size Download all
md5:b54285ddb00ab38d8e57f3791b4dc716
675.2 kB Preview Download

Additional details

Dates

Updated
2026-05-05
v16.0 – Reframed as pre-standardization draft throughout (title, abstract, foreword, submission note, conclusion, cover memo). Em-dash sweep: 20 prose instances replaced with contextually appropriate alternatives (colons, semicolons, parentheses, commas). FedRAMP language softened to "intended to support future mapping." Section 3.1 heading changed from "The Model Is a Commodity" to "Behavioral Guarantees Must Be Portable Across Model Providers." Conclusion slogan removed; replaced with standards-facing close. Abstract corrected from "production-equivalent environment" to "development-laboratory environment with live operational use." Section 6.2 comma-splice corrected. All changes based on external peer review (two independent reviewers).