Framework 01 of 12 · Pre-delegation Phase · Mission Quality
Intent Architecture Framework
The design of the pre-execution contract between human and agent
Every agentic experience begins with a moment of translation. A human has something they want to achieve - a financial goal, a purchase decision, a healthcare outcome - and they must communicate that intention to a system that will act autonomously on their behalf. The quality of that translation determines everything that follows. A poorly specified intent produces a perfectly executed wrong outcome. A well-specified intent enables the agent to act with confidence, within boundaries, toward results the human actually wanted.
The Intent Architecture Framework addresses this foundational challenge. It sits before delegation in the agentic experience lifecycle - the moment before authority is granted, when the human and agent negotiate what "success" actually means. Traditional interface design assumed the user would be present throughout the interaction, correcting course in real time. In agentic systems, the user is absent during execution. The intent specification is the only design input the agent has to work with. If it is incomplete, ambiguous, or misunderstood, there is no interface for the human to correct course mid-flight.
This framework is not about natural language processing or prompt engineering. It is about the design of structured conversations between humans and agents that produce high-fidelity intent objects - complete specifications of goals, constraints, success criteria, exceptions, and what must never happen. It is the architecture of the pre-execution contract that governs everything the agent will do.
Six principles that govern intent specification
These principles define how intent should be captured, structured, and validated before any agent acts. They are drawn from the founding principles of Agentic Experience Design and stress-tested against real-world deployment in financial services and healthcare.
Goals Over Instructions
Intent Architecture distinguishes between goals and instructions. A goal is an outcome the human wants to achieve - 'find me the best mortgage rate for my circumstances.' An instruction is a step-by-step directive - 'go to these three comparison sites and sort by APR.' Agentic systems must elicit goals, not instructions, because the agent may discover better paths to the outcome than the human could have specified. The framework designs for goal elicitation as the primary interaction pattern, with instruction-level detail available as an optional constraint layer for users who want tighter control.
Constraints Are First-Class Citizens
In traditional UX, constraints are edge cases handled by validation rules. In Intent Architecture, constraints are as important as the goal itself. A constraint defines what must not happen, what the agent must never do, and what boundaries must be respected even if they reduce the optimality of the outcome. 'Find me the cheapest flight' is a goal. 'Never book a connection through an airport I have not approved' is a constraint. The framework treats constraints as first-class design objects with their own elicitation, encoding, and verification patterns.
Ambiguity Is a Design Problem, Not a User Failure
When a human says 'get me something nice for dinner,' the ambiguity is not a failure of expression - it is a natural feature of human communication. Intent Architecture designs for ambiguity negotiation: structured dialogues where the agent surfaces its interpretation, identifies gaps, and proposes clarifying questions without overwhelming the user. The framework defines when to ask, when to infer, and when to present a plan for approval rather than seeking perfect specification upfront.
Success Criteria Must Be Explicit
An agent cannot evaluate its own performance without knowing what success looks like. Intent Architecture requires that every delegation includes explicit success criteria - measurable, observable conditions that define when the mission is complete. These criteria serve dual purposes: they guide the agent's decision-making during execution, and they provide the human with a clear basis for evaluating the outcome after the fact. Without explicit success criteria, the agent optimises for its own interpretation of 'good enough.'
Plan Preview Before Execution
Before an agent acts, the human should see what it intends to do. The Plan Preview is a standard pre-execution artefact in Intent Architecture - a structured summary of the agent's proposed approach, including the steps it will take, the decisions it will make autonomously, the points where it will check back, and the constraints it will respect. This is not a confirmation dialog. It is a negotiation surface where the human can adjust scope, tighten constraints, or grant additional authority before execution begins.
Intent Objects Are Persistent and Auditable
The intent specification is not a transient input that disappears after the agent begins acting. It is a persistent, auditable artefact that remains available throughout the agent's operation and after completion. The human can review what they asked for, the agent can reference it when making decisions, and the system can use it for post-execution evaluation. In regulated industries, the intent object becomes part of the compliance record - evidence that the agent acted within the scope of what was requested.
The most consequential design decision in any agentic system is not how the agent acts - it is how the human specifies what the agent should achieve. Intent Architecture is the discipline of making that specification as rich, precise, and auditable as the outcome it governs.
Five patterns for intent specification
Goal Elicitation Design
The structured conversation pattern that helps humans articulate what they want to achieve rather than how they want to achieve it. Uses progressive disclosure to move from high-level goals to specific parameters without overwhelming the user with upfront complexity.
When to use: At the start of every delegation, before any authority is granted or action is taken.
Ambiguity Negotiation Interfaces
Interactive surfaces where the agent presents its interpretation of ambiguous intent and proposes clarifying options. Designed to resolve uncertainty without making the user feel interrogated. Includes confidence indicators showing how well the agent believes it understands the request.
When to use: When the agent's confidence in its interpretation falls below the threshold for autonomous action.
Constraint Encoding
A structured format for capturing scope boundaries, budget limits, reversibility requirements, and absolute prohibitions. Constraints are encoded as machine-readable objects that the agent references throughout execution, not as natural language annotations that might be misinterpreted.
When to use: During intent specification for any delegation involving financial transactions, irreversible actions, or regulated domains.
Success Criteria Specification
A pattern for defining measurable, observable conditions that constitute mission completion. Includes templates for common goal types (purchase, research, scheduling) and custom criteria builders for novel delegations. Success criteria are validated against the goal to ensure they are achievable and complete.
When to use: As a required step in every delegation flow, immediately after goal elicitation and constraint encoding.
Plan Preview Artefact
A pre-execution summary that shows the agent's proposed approach, decision points, check-back triggers, and constraint adherence plan. Serves as the final negotiation surface before the human grants execution authority. The plan preview is persistent and becomes part of the audit trail.
When to use: Before execution begins on any delegation with consequence above the user's auto-approve threshold.
Intent Architecture in agentic commerce
In agentic commerce, the machine customer acts on behalf of the human consumer. The quality of the purchase intent specification directly determines whether the agent buys what the human actually wanted. These are the primary commerce scenarios where Intent Architecture applies.
Purchase Goal Specification
When a consumer delegates a purchase to an agentic shopping agent, the intent must capture not just 'what' but 'why.' A request for 'a new laptop' is insufficient. The Intent Architecture patterns elicit the use case (gaming, business travel, creative work), the constraints (budget ceiling, brand preferences, delivery timeline), and the success criteria (performance benchmarks, weight limits, must-have features). The agent then searches, compares, and negotiates within these parameters - acting as a machine customer with a clear mandate.
Financial Service Delegation
In banking and investment, intent specification is a regulatory requirement as much as a design challenge. When a customer delegates portfolio rebalancing to an agentic AI system, the intent object must capture risk tolerance, ethical investment constraints, liquidity requirements, and tax optimisation preferences. The Plan Preview pattern becomes a regulatory artefact - evidence that the customer understood and approved the agent's proposed strategy before execution.
Healthcare Appointment Orchestration
A patient delegating appointment scheduling to an agent must specify not just 'book me a dermatologist' but the full context: urgency level, geographic constraints, insurance compatibility, preferred times, and whether they need a specialist referral first. The Ambiguity Negotiation pattern is critical here - the agent must surface medical context questions without practising medicine, and the constraint encoding must respect both patient preferences and clinical protocols.
Supply Chain Procurement
Enterprise procurement agents act on behalf of organisations, not individuals. Intent Architecture in this context must capture organisational policies, approved vendor lists, sustainability requirements, and multi-stakeholder approval workflows. The intent object becomes a procurement mandate - a structured specification that multiple agents in the supply chain can reference when negotiating, ordering, and verifying deliveries.
In agentic commerce, the machine customer cannot ask the human what they meant mid-transaction. The intent object is the only contract. If it is incomplete, the agent will optimise for the wrong outcome - and the human will not know until the result arrives.
Intent Architecture Framework: Guidance for Teams
Start With
- -Map the top 5 delegation types your users will perform
- -Define the minimum viable intent object for each delegation type
- -Build a goal elicitation flow for your highest-frequency use case
- -Create constraint templates for your domain's regulatory requirements
Build Toward
- -Adaptive elicitation that learns from user patterns over time
- -Cross-session intent continuity - remembering preferences from previous delegations
- -Multi-stakeholder intent aggregation for enterprise use cases
- -Intent quality scoring that predicts outcome alignment before execution
Measure By
- -Intent-to-outcome alignment rate - did the agent achieve what the human actually wanted?
- -Ambiguity resolution efficiency - how many clarifying rounds before execution?
- -Constraint violation rate - how often does the agent breach specified boundaries?
- -Plan Preview acceptance rate - how often do users approve the agent's proposed approach without modification?
Intent Architecture Framework: Lifecycle Connections
Framework 02
Delegation Design Framework
Intent Architecture defines what the human wants; Delegation Design defines the authority granted to pursue it.
Explore frameworkFramework 11
Onboarding & Capability Discovery
Users cannot specify intent well without understanding agent capabilities. Onboarding builds the mental model that Intent Architecture depends on.
Explore frameworkFramework 03
Autonomy Gradient Design System
The quality of intent specification determines how much autonomy can safely be granted. Poor intent requires tighter autonomy constraints.
Explore frameworkIntent Architecture Framework: What Comes Next
Intent Architecture defines what the human wants. The next framework in the lifecycle - Delegation Design - defines how authority is granted to pursue it. Together, they form the pre-execution foundation of every agentic experience.
Intent Architecture Framework: The Framework Ecosystem
Navigate the complete lifecycle of Agentic Experience Design. Each framework addresses a distinct phase of the human-agent relationship.