Agent Reliability Profile

How do you know an AI agent will stay in its lane? The Profile gives institutions, vendors, assurance firms, and regulators a standardized way to communicate and test agent reliability. Each Profile states what a deployed agent can do, under which controls, and how strongly those have been tested. The Profile is designed to support a simple, but critical claim:

“This agentic system reliably functions within its operating boundary.”

Working draft circulated for comment arXiv:2610.04123 October 2026

The operating boundary of an agent

An agent’s operating boundary is defined by how much it can do on its own (autonomy), where it is allowed to work (ODD), what it can and can’t do (action classes), and what contains it (control envelope). This four-part taxonomy turns the assertion “This agent is reliable” into a scoped claim that can be tested and falsified. Standardizing enables scale, risk management, and trust.

Autonomy tier (T)

How much it does without a person, and who takes over when it cannot continue.

T0T1T2T3T4U
Five tiers plus U, “unprofiled”: a finding, not a tier. Set by evidence, not by configuration.

Operational design domain (O)

Where it is allowed to work: use case, data, customers, jurisdiction and cost of failure.

ODD.1…ODD.52
52 candidate use cases across nine business lines. Does the agent notice when it has left, and stop?

Action classes (A⁺ A⁻)

What it may do and what it must never do, ranked by how hard each action is to undo.

A1 read…A9 move fundsA10 file…A12 spawn
Twelve classes, stated as a permitted set and a prohibited set.

Control envelope (C)

What contains it: the layered controls that make this tier acceptable in this domain.

modeltoolharnessenvironmentinter-agentcontrol plane
55 named controls across six layers, tested as a whole.

Bounded claim S(D)

reliable at T, within O, for A⁺ (never A⁻), under C. The claim fixes the tests, the pass-or-fail floors and the validity window.

Agent Reliability Profile P(D)

claim + floors + provenance + assurance level + out-of-scope statement + evidence pointers

Profile Package

signed inputs, traces, evaluator outputs, assessor records. Stays with the institution.

Three levels of assurance, one Profile

Each agent gets a Profile, which can be one of three levels. The Profile keeps the same shape at every level. What changes is the degree of testing and independence. Benchmark results are a separate measurement.

Level 1

Asserted Profile

What happens
An institution’s documents and settings for its agent are read and compiled into a Profile. Nothing is executed.
Who stands behind it
The institution
What it may claim
“This is what we believe we have deployed.”
Level 2

Validated Profile

What happens
The institution runs tests against its agent, in its own environment.
Who stands behind it
The institution
What it may claim
“No critical control failure under the conditions we tested.”
Level 3

Verified Profile

What happens
A qualified independent assessor runs tests against an institution’s agent and signs its own report.
Who stands behind it
An independent assurance firm
What it may claim
“No critical control failure, independently tested by someone else.”
Separate field
Benchmark comparability flag · Profile Benchmarker

Benchmarked Profile

An independent assessor runs the agent in a shared reference environment with a hidden, rotating scenario set. It enables comparability across deployments and institutions.

The Profile Kit

Everything an engineering team needs to produce, check and read Profiles, released as an open-source MLCommons repository on GitHub.

The Kit holds the Profile specification, which defines what counts as a valid Profile, and the reference tools that put it into practice. Anyone can build tools that produce a valid Profile, as with MLPerf; where a tool and the specification disagree, the specification wins. The reference tools are designed to run inside your institution, and they stay model-agnostic. The public release covers Level 1 (Asserted) Profiles; Level 2 testing is available in Working Group pilots.

Illustrative only. Field names may change before release.

Taxonomy repository

Draft for comment · v0.1

The shared vocabulary: autonomy tiers, use cases, action classes and controls, each with a permanent code. Published first, for consultation, so supervisors, standards bodies and vendors can cite and build on it.

Creative Commons Attribution 4.0See the Taxonomy

Profile Kit repository

In preparation

The specification and reference tools below. Each Kit release takes its vocabulary from a numbered taxonomy release, and every Profile names the versions it was written against.

Apache 2.0Get the Profile Kit

Profile Specification

Release 1

The rules a valid Profile must meet: required sections, how evidence is labelled, what each assurance level may claim, and the rules for flags. It names the taxonomy release it uses.

Normative · Working Group

Conformance suite

Release 1

Test vectors and a reference checker. Use them to confirm that the Profiles your tool, pipeline or assessor’s harness produces conform to the specification.

Normative · Working Group

Profile Builder skill

Release 1

An agent skill that compiles a Level 1 Profile from your documents and configuration, flags conflicts between them, and leaves unknown fields blank. It runs locally, and your institution controls where evidence goes.

Reference tool

Validator & adapters

Release 2

Runs scenarios against a deployed agent for Level 2, with an evidence event schema and reference adapters so validation does not need custom code for every platform.

Reference tool · contributions welcome

Basic Kit

Release 1

A spreadsheet template, a one-row Profile Lite for agent inventories, and an offline Profile Viewer. For teams that do not run code.

Template

Regulator Kit

Preview

A supervisory console for reading agent inventories and Profiles, with trends, outliers and changes between submissions.

For supervisors

See it in action

Interactive demo

Assurance Console

Built for the C:\>DIR hackathon. Choose a target agent at one of three fictional firms, build its Profile from documents and configuration, review the result, and inspect the verified test runs behind it.

  1. Step 1Select a target
  2. Step 2Build the Profile
  3. Step 3Review the Profile
  4. Step 4Verified runs
3 synthetic firms124 claims28 boundary runs1,617 checks

The demo is password protected and may take up to a minute to wake.

Build the Profile with us

The most durable standards are built on trust and active participation by diverse stakeholders. The Working Group includes leading banks, fintechs, networks, assurance firms, and regulators. Come join us as we build and improve the Profile.

Shape the standard

Help set the vocabulary, the specification rules and what counts as a valid Profile, before supervisors settle on one.

Pilot in a sandbox, with support

Use an MLCommons-hosted reference sandbox with synthetic agents and a scenario library, plus engineering support, under an agreed testing protocol.

Get a head start with the Kit

Work with the specification and reference tools before public release, and contribute adapters and domain packs.

Work alongside supervisors

Regulatory observers join methodology sessions. Each bank decides what, if anything, observers see.

Working Group

Financial Services WG membership

  • Working Group membership
  • Full access to WG tools and content
  • Fee credited toward sponsorship or benchmark use
Sponsor

Full MLCommons membership

  • Voting rights
  • Named sponsor of the AILuminate Financial Services benchmarks
  • Everything in WG membership
Lead sponsor

Industry co-chair

  • All sponsor benefits
  • An industry co-chair seat
  • Limited availability

How to join

  1. Fill out the subscription form and choose the AI Risk & Reliability Working Group.
  2. Associate a Google account with your organizational email address.
  3. Once approved, you get the meeting invitation and access to the Working Group’s shared folder.

For sponsorship and the pilot cohort, please complete this form.

Frequently asked questions

Is a Profile a score, grade or certification?

No. A Profile is a structured claim with evidence and conditions. Metrics are reported separately rather than combined into a score, and pass-or-fail floors cannot be outweighed by good averages. Levels 1 and 2 are self-assurance and are labeled as such on the Profile.

Is it a standard?

Not yet. The preprint is a working draft circulated for comment. To become a standard, the Profile has to be used repeatedly, show predictive value, and be revised in response to failures. The pilot is designed to test exactly that.

Does any of our data leave the institution?

Not at Levels 1 and 2. At Level 3, either the evidence travels to the independent assessor or the assessor works on site. Benchmarking may involve sharing, depending on the components benchmarked. A full Profile can itself contain sensitive details, such as evidence locations, short quotations and configuration values, so the Kit asks for a disclosure review before one is shared.

How does it relate to NIST AI RMF, the EU AI Act, ISO/IEC 42001 and SR 26-2?

Those frameworks call for per-deployment evidence without specifying its form. The Profile supplies it. The preprint includes a crosswalk to NIST AI RMF, the EU AI Act, the CRI Profile and ISO/IEC 42001. Jurisdiction-specific mappings, including SR 26-2, are planned.

What about agents built on our internal agent platform?

The platform can carry its own Platform Baseline Profile: a claim about the ceilings it enforces on every session it hosts. Agents built on it carry ordinary Profiles anchored to the platform’s. A Level 3 platform Profile can cover the substrate while each anchored agent stays at Level 1. Platform Baseline Profiles are described in the preprint but are not part of the current release candidate of the Kit.

Do we have to use MLCommons tools?

No. The Profile is software-neutral. Any tool that produces a Profile meeting the specification is acceptable, whether it comes from a vendor, an assurance firm or your own engineers.

Questions?

For membership, sponsorship, the pilot cohort or demo access, write to the membership team.

Working Group chairs
Mike HsuFellow, Cambridge Judge Business School; former Acting Comptroller of the Currency
Medha BankhwalCo-founder and CEO, Mezuro
Deborah EngAI and data standards and governance, JPMorganChase; Chair, Global Standards Committee, Cyber Risk Institute
Back to the Financial Services Working Group

Privacy Overview

This website uses cookies so that we can provide you with the best user experience possible. Cookie information is stored in your browser and performs functions such as recognising you when you return to our website and helping our team to understand which sections of the website you find most interesting and useful.