Not all agents are the same. The term “AI agent” has become a catch-all for any autonomous system that acts on behalf of a human, but this conflation obscures the critical differences between agent types - differences in function, in delegation pattern, in trust requirement, and in the design challenges they present. A comparison agent that evaluates products across merchants is fundamentally different from a negotiation agent that haggles over price. A budget agent that manages spending constraints operates under different trust requirements than a purchasing agent that executes end-to-end transactions.
This essay proposes a taxonomy of commerce agents - a classification system that maps the primary agent types by their function, their delegation design, and their trust architecture requirements. The taxonomy is not exhaustive. New agent types will emerge as agentic commerce matures. But the five categories identified here - comparison, negotiation, budget, subscription, and purchasing - represent the foundational types from which more specialised agents will be composed.
I. The Comparison Agent
The comparison agent is the most constrained and the most common type of commerce agent. Its function is evaluation: given a set of criteria from the human principal, the comparison agent surveys available merchants, products, or services and returns a ranked assessment. It does not purchase. It does not negotiate. It does not commit the principal to any transaction. It evaluates and recommends.
The delegation pattern for a comparison agent is narrow but information-intensive. The principal specifies: what to compare (product category, service type), the criteria for comparison (price, quality, delivery speed, return policy, sustainability credentials), the weighting of those criteria (price matters more than delivery speed, or vice versa), and the scope of the comparison (specific merchants, or the entire market). The agent operates within these constraints, processing structured data from merchants and returning a recommendation.
The trust requirement for a comparison agent is primarily external - it must trust the data it evaluates. A comparison agent that cannot verify whether a merchant’s price is current, whether the product specifications are accurate, or whether the return policy is as stated cannot fulfil its mandate. This makes the comparison agent the most sensitive to merchant data quality. Merchants with structured, verified, machine-readable product data are visible to comparison agents. Merchants with unstructured, unverified, or human-only data are invisible.
The comparison agent also faces a trust challenge that is unique to its function: the risk of biased evaluation. If the comparison agent is provided by a platform that also sells advertising or preferred placement, the agent’s recommendations may be influenced by commercial relationships rather than the principal’s criteria. Delegation design for comparison agents must address this conflict of interest explicitly - the principal must be able to verify that the agent’s evaluation was based solely on the specified criteria, not on undisclosed commercial incentives.
II. The Negotiation Agent
The negotiation agent is the most complex type of commerce agent. Its function is not merely to evaluate or execute, but to engage in dynamic interaction with merchants to achieve better terms for the human principal. Negotiation requires strategic reasoning: when to concede, when to hold firm, when to walk away, when to accept. It requires the ability to make commitments - “my principal will purchase at this price” - that are binding within the scope of the delegation.
The delegation pattern for a negotiation agent is the most detailed of any agent type. The principal must specify: the target price or terms, the acceptable range (the worst terms the principal will accept), the concessions the agent is authorised to make (free shipping in exchange for a higher price, longer delivery window in exchange for a discount), the terms that are non-negotiable (minimum quality, maximum delivery time), and the conditions under which the agent must escalate to the human rather than continuing to negotiate autonomously.
The trust architecture for negotiation agents must address a problem that does not arise with other agent types: the enforceability of agent-negotiated agreements. When a human negotiates with a merchant, the agreement is between two legal persons. When an agent negotiates with a merchant, the agreement is between a machine and a legal person, on behalf of another legal person who may not be present. Is the agreement binding? Under what conditions can the principal repudiate an agreement their agent made? What happens if the agent exceeded its negotiation authority?
These questions are not theoretical. As negotiation agents become more capable, they will negotiate prices, delivery terms, warranty conditions, and service level agreements on behalf of human principals. The trust architecture must ensure that the merchant knows the agent’s authority is genuine, that the principal knows the agent is negotiating within its mandate, and that both parties can verify the terms of the agreement after the fact. Verifiable intent - the cryptographic proof that the agent’s actions are authorised by the principal - becomes essential for negotiation agents.
III. The Budget Agent
The budget agent manages the financial dimension of the human principal’s commerce activity. Its function is governance: ensuring that the principal’s spending remains within defined constraints across multiple transactions, multiple merchants, and multiple time periods. The budget agent does not select products or negotiate prices - it constrains the actions of other agents (or of the human directly) by enforcing financial boundaries.
The delegation pattern for a budget agent is financial and temporal. The principal specifies: the total budget for a period (weekly, monthly, quarterly), the allocation across categories (groceries, entertainment, subscriptions, discretionary), the rules for reallocation (can unused grocery budget be spent on entertainment?), and the escalation thresholds (notify me when spending reaches 80% of budget, block transactions above 90%). The budget agent maintains a running ledger of expenditure and enforces these constraints in real time.
The trust requirement for a budget agent is primarily internal - the principal must trust that the agent accurately tracks spending and enforces constraints without exception. But the budget agent also requires external trust: it must trust merchant pricing data to prevent overspending due to inaccurate or manipulated prices. If a merchant reports a price of £10 but charges £15, the budget agent’s ledger is incorrect and its constraints are violated. The budget agent must verify transaction amounts against the merchant’s stated prices and flag discrepancies immediately.
The budget agent is the financial governor of the principal’s agentic commerce activity. It sits above other agent types in the authority hierarchy - a purchasing agent cannot execute a transaction that the budget agent has not approved. This hierarchical relationship between agents is a design challenge that the AXD Institute addresses through the concept of delegation scope: each agent operates within a defined scope of authority, and the budget agent’s scope encompasses the financial dimension of all other agents’ activities.
IV. The Subscription Agent
The subscription agent manages ongoing purchasing relationships on behalf of the human principal. Its function is maintenance: ensuring that recurring purchases continue to serve the principal’s interests, that subscription terms remain favourable, and that alternatives are evaluated at each renewal cycle. The subscription agent is the operational expression of delegation maintenance - the continuous work of keeping a delegated relationship healthy.
The delegation pattern for a subscription agent is temporal and evaluative. The principal specifies: which subscriptions to manage, the criteria for continued renewal (price threshold, quality threshold, usage threshold), the conditions for cancellation (price increase above a percentage, quality degradation, better alternative available), and the notification preferences (inform me before cancelling, or cancel autonomously within these parameters). The subscription agent evaluates each subscription at each renewal cycle against these criteria.
The trust requirement for a subscription agent is bidirectional. The agent must trust the merchant’s subscription data - usage metrics, price change notifications, service quality indicators - to make informed renewal decisions. And the merchant must trust that the agent represents a genuine human principal with valid payment credentials. The subscription agent operates in a continuous relationship with the merchant, unlike the purchasing agent which may interact with a merchant only once. This continuity creates opportunities for trust accumulation but also for trust degradation if the merchant’s service quality declines over time.
The subscription agent is particularly important in the agentic economy because subscriptions represent the largest category of recurring autonomous transactions. Software licences, streaming services, grocery deliveries, insurance renewals, utility contracts - all of these are subscription relationships that agents will manage on behalf of human principals. The design of subscription agents will determine whether the subscription economy becomes more transparent and competitive (because agents continuously evaluate alternatives) or more concentrated (because agents default to established relationships to minimise evaluation cost).
V. The Purchasing Agent
The purchasing agent is the most autonomous type of commerce agent. Its function is end-to-end transaction execution: receiving a mandate from the human principal, evaluating merchants, selecting a product, negotiating terms (if authorised), executing payment, and monitoring fulfilment. The purchasing agent combines elements of the comparison agent, the negotiation agent, and the budget agent into a single autonomous system.
The delegation pattern for a purchasing agent is the broadest of any agent type. The principal specifies: what to purchase (product category, specific requirements), the constraints (maximum price, minimum quality, delivery requirements), the payment method, and the conditions for autonomous execution versus human escalation. The purchasing agent operates within these constraints but has significant discretion in how it fulfils the mandate - which merchant to select, which specific product to choose, what terms to accept.
The trust requirement for a purchasing agent is the highest of any agent type, because it makes binding financial commitments on behalf of the human principal. The principal must trust that the agent will act within its mandate, select merchants that are trustworthy, and execute transactions that serve the principal’s interests. The merchant must trust that the agent represents a genuine principal with valid authority and payment credentials. The payment provider must trust that the transaction is authorised. Every actor in the transaction chain must trust every other actor - and the purchasing agent sits at the centre of this trust web.
The purchasing agent is the machine customer in its fullest expression. It is the agent that actually transacts - that moves money from the principal’s account to the merchant’s account in exchange for goods or services. The design of the purchasing agent’s delegation, trust architecture, and operational envelope is the central challenge of Agentic Experience Design. Every other agent type supports, constrains, or informs the purchasing agent’s decisions. The purchasing agent is where the entire AXD framework converges.
VI. Cross-Cutting Design Principles
Across all five agent types, several design principles apply universally. First, delegation must be explicit. Every agent must operate within a clearly defined scope of authority that the human principal has deliberately granted. Implicit delegation - where the agent assumes authority that was not explicitly granted - is a trust violation regardless of agent type.
Second, trust must be earned, not assumed. No agent should trust a merchant’s data without verification. No merchant should trust an agent’s authority without credential checking. No human should trust an agent’s judgment without transparency into its decision-making process. Trust is the primary material of every agent type, and it must be designed, not presumed.
Third, escalation must be designed. Every agent type encounters situations that exceed its delegation scope or its confidence threshold. The design of escalation - when to involve the human, how to present the situation, what information to provide - is as important as the design of autonomous operation. An agent that escalates well maintains trust. An agent that fails to escalate when it should, or that escalates too frequently, erodes trust.
Fourth, agents compose. In practice, a human principal will not deploy a single agent type in isolation. They will deploy a purchasing agent that consults a comparison agent, that is constrained by a budget agent, that manages subscriptions through a subscription agent. The design of agent composition - how agents communicate, how authority flows between them, how conflicts are resolved - is a design challenge that the taxonomy makes visible but does not resolve. The composition of agents is the next frontier of Agentic Experience Design.