What technical evidence may show
System records may show which files were accessed and what outputs or actions occurred, depending on the implementation and its logging.
OPERATIONAL FRAMEWORKS
The three organizational layers every enterprise must design before any agent goes live. Containment is Layer 1. Almost no enterprise builds Layer 3.
Intent documented at deployment is not the same as intent maintained over time. The stack was built around that distinction because most governance failures are not authorization failures at launch. They are drift failures six months later when the original decision-makers are gone.
Published under CC BY 4.0. Free to reproduce, adapt, translate, and use commercially, including inside your own governance program, with attribution to Sougata Roy and a link to this page. Attribution is a condition of the license. Claiming authorship is not attribution. Full terms at sougataroy.com/rights
Cite this framework
Sougata Roy, "The Intent Architecture Stack", Version 1.0, April 2026, https://sougataroy.com/frameworks/intent-architecture-stack
At-a-glance index. Open any card marked “Read the detail” to see its supporting explanation.
Layer 1
Layer 1 establishes the organizational context within which the agent will operate. This layer must be completed before Layer 2 is written. The context determines what the intent can permissibly be.
Layer 2
Layer 2 is the organizational record of what the agent was built to do, expressed in plain language that a compliance officer, a regulator, or a board member can evaluate without technical context. It is derived from the Context Document produced in Layer 1.
Layer 3
Layer 3 establishes the organizational accountability structure that governs the agent throughout its operational life. This is where the Intent Architecture Stack operationalizes what NIST AI RMF's GOVERN function, EU AI Act Article 26, and the CSA Agentic Profile require at the agent level.
Version 1.0 - Published April 2026
THE IDEA IN 30 SECONDS · CONSTRUCTED EXAMPLE
A fictional legal-operations team is preparing an agent to identify non-standard contract clauses. The example follows the existing framework's completed example; it does not represent a real deployment.
The three-layer decision path
Name affected stakeholders, data, obligations, and connected systems.
Status: What must the organization account for?
Illustrative part 1 of 3.Identify clause variance and route it to Legal Operations for review.
Status: No approval, rejection, editing, or external sending.
Illustrative part 2 of 3.Name the consequence owner, review cadence, and escalation path.
Status: Revisit when scope, data, regulation, or ownership changes.
Illustrative part 3 of 3.System records may show which files were accessed and what outputs or actions occurred, depending on the implementation and its logging.
The business owner approves the purpose and action boundary. The technical design must then route or prevent actions that exceed it.
Type
Three-layer governance framework
Version
v1.0
Published
April 2026
Time to use
20 min review / 60 min workshop
Audience
Business, risk, security, and compliance leaders
Output
A signed authorization record
Use this first
The three organizational layers every enterprise must design before any agent goes live. Containment is Layer 1. Almost no enterprise builds Layer 3.
Primary object
Read top to bottom

Three-layer stack
The stack becomes complete only when the context document, intent document, and governance record form one traceable authorization record.
Limitation
This object can prove what the organization documented, who accepted the decision, and whether the three layers are present. It cannot prove that production behavior followed the record, that technical permissions enforce every boundary, or that the named owner responded effectively to an incident.
Revision history
August 3, 2026: Page structure reordered to place the primary artifact before explanatory content; constructed completed example added. Framework definition and version unchanged; remains v1.0. v1.0, April 2026: Original publication. The framework was subsequently developed in the white paper Who Owns the Agent? (May 2026).
Copyable citation
Sougata Roy, "The Intent Architecture Stack," Version 1.0, April 2026, https://sougataroy.com/frameworks/intent-architecture-stack
The problem
Intent Architecture exists because most agent deployments can show what the platform allowed, but cannot show what the organization formally decided.
Use this section to locate the failure: the missing record that says what the agent was supposed to do before anyone reads the audit log.
Most organizations that deploy AI agents spend significant effort on what the agent can do technically and almost no effort on what the organization has formally decided it should do. Those are different questions. The first is answered by the platform. The second is answered by the organization, or not answered at all.
The gap between those two questions is where governance failures live. The Intent Architecture Stack produces the authorization record before deployment. Without that record there is no baseline against which production behavior can later be compared, which is what makes the Intent Gap unmeasurable rather than merely open. When an AI agent produces a harmful output, the audit log shows what happened. It does not show what was supposed to happen. Without an authorization record, the log cannot be evaluated against any standard. The Intent Architecture Stack is the organizational design that creates that standard before deployment, not after something goes wrong.
Governance question
Before an AI agent is deployed in your environment, does your organization have a formal record of what it was authorized to do, who made that authorization, and under what conditions that authorization must be reviewed?
Book canon note
Relationship to the book. The Intent Architecture Stack originates in the white paper Who Owns the Agent? (May 2026) and is not among the defined terms in the book of the same name. The book's vocabulary covers the authorization decisions the stack produces rather than the stack itself.

Intent gap
The framework starts by separating what the platform allowed from what the organization formally approved before deployment.
Constructed example. Agent IDs, financial resources, authorization records, dates, and audit entries are illustrative values, not real records.
The stack
The Intent Architecture Stack has three organizational layers. Each one is a governance decision, not a technical configuration. All three must exist before an agent is considered authorized to operate. Most organizations build Layer 1 informally, build Layer 2 partially, and skip Layer 3 entirely until something forces the question.
Read from Layer 1 to Layer 3. Context defines what is permissible, intent defines what is authorized, and governance defines who answers for it.

Three-layer stack
The stack becomes complete only when the context document, intent document, and governance record form one traceable authorization record.
Layer 1 establishes the organizational context within which the agent will operate. This layer must be completed before Layer 2 is written. The context determines what the intent can permissibly be.
Accountability test
Layer 1 is complete when the Context Document exists as a pre-deployment artifact, not as a post-hoc description.
Layer 2 is the organizational record of what the agent was built to do, expressed in plain language that a compliance officer, a regulator, or a board member can evaluate without technical context. It is derived from the Context Document produced in Layer 1.
Accountability test
The test for Layer 2: can any stakeholder in your organization answer 'what is this agent authorized to do and who approved it' without consulting the original developers? If no, Layer 2 does not exist for that agent.
Layer 3 establishes the organizational accountability structure that governs the agent throughout its operational life. This is where the Intent Architecture Stack operationalizes what NIST AI RMF's GOVERN function, EU AI Act Article 26, and the CSA Agentic Profile require at the agent level.
Accountability test
Layer 3 is complete only when all three sub-components are documented as pre-deployment artifacts and the Consequence Owner has confirmed their role in writing.
Layer breakdown
The stack is not a metaphor. Each layer has named artifacts, required fields, and a test for whether the organization has actually built it.
Use the breakdown as a pre-deployment review sequence. A missing field is a governance finding, not a formatting issue.
Layer 1
Layer 1 establishes the organizational context within which the agent will operate. This layer must be completed before Layer 2 is written. The context determines what the intent can permissibly be.
Use this layer as a document test: if the organization cannot produce the artifact named here before deployment, the layer is not complete.

Layer 1 visual
The environment, affected data, stakeholder impact, and downstream integrations set the boundary before any purpose statement is written.
REGULATORY ENVIRONMENT
Define regulatory obligations before defining intent. Map the regulatory frameworks applicable to your industry, identify specific obligations triggered by the agent's data access and decision scope, and document which regulatory bodies have examination authority over the deployment context.
STAKEHOLDERS AND DATA
Identify every stakeholder group affected by the agent's outputs and every data touchpoint the agent can reach. This includes upstream data sources the agent reads, downstream systems the agent can trigger, and individuals or groups whose data the agent processes.
SYSTEM INTEGRATIONS
Map all downstream system triggers and integration points to define the full technical scope. An agent that appears limited by its interface may still have significant reach through the systems it can invoke. The scope of system integration determines the outer boundary that Layer 2 must address.
Present when
Organization can produce a Context Document, prepared before the Intent Document was written, naming applicable regulatory obligations, affected stakeholder groups, and every downstream system the agent can trigger or access.
Absent when
Regulatory environment, stakeholder impact, and system scope were assumed rather than documented before intent was defined.
Control point
Layer 1 is complete when the Context Document exists as a pre-deployment artifact, not as a post-hoc description.
Layer 2
Layer 2 is the organizational record of what the agent was built to do, expressed in plain language that a compliance officer, a regulator, or a board member can evaluate without technical context. It is derived from the Context Document produced in Layer 1.
Use this layer as a document test: if the organization cannot produce the artifact named here before deployment, the layer is not complete.

Layer 2 visual
The intent document names the agent purpose, scope, prohibitions, expected outputs, and review triggers in plain language.
Constructed scenario. The finance workflow, systems, scope, and review conditions are examples, not an authorization for a real agent.
PURPOSE STATEMENT
Write the agent's purpose in one plain-language sentence that states what problem the agent solves and what workflow it supports. Not a technical description. Not a feature list. A statement of organizationally authorized intent: 'This agent reviews contract documents for non-standard clauses and flags them for legal review before signature.'
AUTHORIZED SCOPE
Define explicitly what the agent is permitted to do and what it is explicitly prohibited from doing. The prohibitions field is not optional. An authorization record without explicit prohibitions is incomplete. The scope must be expressed in terms of actions, data access, and system boundaries - not in terms of model capabilities.
EXPECTED OUTPUTS
Specify what correct agent behavior looks like and define the triggers for human review. Expected output format, expected output frequency, and the conditions under which the output must be reviewed by a human before being acted upon.
Present when
A written Intent Document exists with all three components - Purpose Statement, Authorized Scope with explicit prohibitions, and Expected Outputs - prepared before production deployment.
Absent when
The agent's intent is implied by its configuration with no pre-deployment written statement.
Control point
The test for Layer 2: can any stakeholder in your organization answer 'what is this agent authorized to do and who approved it' without consulting the original developers? If no, Layer 2 does not exist for that agent.
Layer 3
Layer 3 establishes the organizational accountability structure that governs the agent throughout its operational life. This is where the Intent Architecture Stack operationalizes what NIST AI RMF's GOVERN function, EU AI Act Article 26, and the CSA Agentic Profile require at the agent level.
Use this layer as a document test: if the organization cannot produce the artifact named here before deployment, the layer is not complete.

Layer 3 visual
Layer 3 is complete only when ownership, review cadence, escalation path, and owner confirmation exist before production use.
Constructed example. The named owner, role, signature, dates, review record, and production approval are fictional sample values.
ACCOUNTABLE OWNER
Assign a named Consequence Owner responsible for board-level accountability and incident escalation decisions. The Consequence Owner is nested within the organization's existing formal accountability structure - three-lines-of-defense, senior management function, risk committee, or equivalent. This person is accountable for governance decisions, not necessarily first-line incident response. The role maps to the Sponsor field in Microsoft Entra Agent ID but is not satisfied by populating that field alone. The organizational accountability structure must be documented separately.
REVIEW CADENCE
Define a recurring review schedule and a set of event-based triggers that require re-authorization. The calendar cadence must be defined before deployment. Event triggers include: change in the agent's data access scope, change in applicable regulatory requirements, change in the named Consequence Owner, and any security incident involving the agent. An authorization record without a review date is an artifact, not governance.
ESCALATION PATH
Write the escalation path before the first incident. The path must name specific contacts and sequences, specify the conditions that trigger escalation at each level, and be executable by someone other than the original developer. If the Consequence Owner cannot describe how to stop the agent and who to call without consulting the builder, the Escalation Path does not exist.
Present when
A named Consequence Owner with board-level accountability nested in formal accountability structure, a defined Review Cadence with calendar date and event triggers, and a written Escalation Path with named contacts - all documented before production deployment.
Absent when
Sponsor fields are populated in the platform but no written organizational accountability structure exists outside the platform configuration.
Control point
Layer 3 is complete only when all three sub-components are documented as pre-deployment artifacts and the Consequence Owner has confirmed their role in writing.
Template download
The full three-layer documentation template covering Context, Intent, and Governance. Complete one per agent before deployment.
Intent Architecture Stack Documentation Template · v1.0 · May 2026 · Sougata Roy, sougataroy.com
Diagnostics
These companion diagnostics apply the framework to specific governance failure patterns.
Use these diagnostics when you need a narrower control map tied back to the stack.
Primary sources
The framework is grounded in the current shift from platform capability to organizational accountability for agent behavior.
Use these references to support the governance case internally and to connect the framework to external authority.
nccoe.nist.gov
Source: NIST National Cybersecurity Center of Excellence, "Accelerating the Adoption of Software and AI Agent Identity and Authorization," February 5, 2026.
View sourcepaloaltonetworks.com/cyberpedia/what-is-agentic-ai-governance
Source: Palo Alto Networks, "A Complete Guide to Agentic AI Governance," February 24, 2026.
View sourcetechcommunity.microsoft.com
Source: Microsoft Tech Community, "Governing AI Agent Behavior: Aligning User, Developer, Role, and Organizational Intent," March 2026.
View sourceConnected frameworks
Intent Architecture gives the authorization record its structure. These adjacent frameworks extend that record into ownership, lifecycle coverage, and orchestration.
Use these cards when the question moves from a single agent's authorization record to ownership, change over time, or multi-agent delegation.
Connected framework
Maps who owns what before deployment, so Layer 3 accountability is not inferred after the agent goes live.
Open frameworkConnected framework
Explains how authorization coverage changes over time as agents, owners, and deployment scope shift.
Open frameworkConnected framework
Extends the authorization question to multi-agent orchestration, delegation, and chained system action.
Open framework