The Fifth Domain

Enterprise Agent Architecture (EAA) treats the agent workforce as enterprise architecture's fifth domain, alongside applications, data, infrastructure, and integration. It organizes governance into four layers — the Agent / Workforce layer, the Capability / Tool layer, the Control Plane, and the Governance layer — and adds the question existing control planes leave out: not just who an agent is and what it may do, but why a given action should proceed.

Enterprises already govern applications, data, infrastructure, and integration. The agent workforce is the domain they have not yet designed for. Agents act continuously, compose tools, and carry state across sessions — yet most control planes govern them the way they govern human users.

A control plane answers two questions well: WHO an agent is (identity) and HOW it may act (permissions and tools). It rarely answers the third: WHY — whether an authorized action is sound right now, and whether a run of individually-authorized actions stays sound across a trajectory.

EAA makes that WHY layer explicit. What follows is a characterization of how to govern autonomous agents — a reference model, not a certification of any one deployment.

Agent / Workforce layer — the agents themselves: their roles, autonomy level, and lifecycle
Capability / Tool layer — the tools, data, and actions each agent can reach
Control Plane — identity, permissions, and routing: the WHO and the HOW
Governance layer — evaluation of WHY an authorized action should proceed, now and over time

WHO, HOW, WHY

Every agent action passes three checks. WHO establishes identity. HOW establishes permitted tools and scope. WHY — the core of the method — evaluates whether an authorized action is sound at this moment and as part of an ongoing trajectory.

1

WHO — Identity

Which agent is acting

Establishes the acting agent's identity, role, and autonomy level before anything runs. This is the layer today's control planes already handle well.

Agent identity and role binding
Autonomy level (who may act unsupervised)
Provenance of the request
Session and lifecycle context

Handled by: the control plane.

2

HOW — Permissions & Tools

What the agent may reach

Establishes which tools, data, and actions the identity is permitted to use, and with what scope. Also control-plane territory — necessary, but not sufficient.

Tool and capability grants
Data and action scope
Least-privilege boundaries
Rate and cost limits

Handled by: the control plane.

3

WHY — Governance

Whether an authorized action is sound

The method's core. An action can be fully authorized and still be unwise — because the situation has changed, or because it is the tenth risky-but-permitted step in a row. The WHY layer evaluates soundness at the moment of action and across the trajectory.

It runs a six-gate evaluation, resolves the result into a single system posture, enforces hard constraints no gate can override, and accumulates risk across sessions so that drift a per-action check would miss becomes visible.

Six-gate evaluation (epistemic, risk, governance, economic, autonomy, constitutional)
System posture: COMPOUND, RUN, THROTTLE, FREEZE, or STOP
Hard constraints — absolute prohibitions no gate result can override
Cross-session risk composition — accumulated risk crossing a budget forces a HOLD, even when each step was individually allowed
A correlated evidence chain linking each decision to its inputs, for accountability

Handled by: the governance layer — the part EAA adds.

What This Method Does NOT Claim

Certification, audit-readiness, or production-hardening of any deployment
A guarantee that a governed agent cannot fail
NIST validation or endorsement — see the note on public-comment submissions below
A replacement for a control plane — it complements identity and permissions, it does not replace them
A finished commercial product — this is an openly-inspectable reference model

What the Governance Layer Produces

A system posture (COMPOUND / RUN / THROTTLE / FREEZE / STOP) for every evaluated action
A HOLD when cross-session risk crosses its budget, even on individually-authorized steps
A correlated evidence chain — inputs, gates, and outcome — for after-the-fact accountability
Enforcement of hard constraints that no gate result can override

Why This Matters

We are not shipping a finished product. We are proving a method in the open, against our own operation.

HRAO-E — the system serving this page — runs the governance layer against itself. It is the live reference implementation, not a demo.

Three questions drive the work: Does an explicit WHY layer change agent behavior in practice? Does cross-session composition catch drift that per-action checks miss? Is the method sound enough for others to adopt and inspect?

Where the evidence does not support a claim, we say so and publish it. That is the commitment.

Published Research

Enterprise Agent Architecture: Governing the Agent Workforce as Enterprise Architecture's Fifth Domain

Saleme, M.K. (2026). Zenodo. DOI: 10.5281/zenodo.21105314

Authorized but Composed: Cross-Session Risk Composition as an Agent-Governance Control

Saleme, M.K. (2026). Zenodo. DOI: 10.5281/zenodo.21400261

Authorized but Refused: Six Months of Runtime Governance Telemetry from an Autonomous Enterprise

Saleme, M.K. (2026). Zenodo. DOI: 10.5281/zenodo.21263262

The Open Reference

constitutional-agent (PyPI, v0.8.0) — the open-core reference implementation of the governance layer. Roughly 330 downloads per month; inspectable line by line.

The adversarial harness — 604 tests across 43 modules — attacks the governance layer to characterize where it holds and where it does not. Security proof, not a security guarantee.

Live composition telemetry — 30,496 refusals across 451,163 evaluated events over roughly six months of operation.

On standards: we have made three public-comment submissions concerning NIST AI 800-2; CAISI acknowledged receipt. That is engagement — not validation, endorsement, or certification.

The Series & Field Notes

We publish the reasoning as it develops:

The Enterprise Agent Architecture framework — the position-paper series and the full reference model. Field notes capture what governing an agent workforce looks like in practice.