Agentic Commerce

Commerce where AI agents act on behalf of users to discover, evaluate, negotiate, and purchase goods/services — replacing or augmenting human decision-making in procurement.

Core Thesis

Traditional commerce relies on information matchmakers (search engines, marketplaces, comparison sites, review aggregators) to connect buyers with sellers. Each layer extracts rent from human attention scarcity.

Agent era disrupts this: Agents collect info directly, compare at scale, and don't need "discovery" in the human sense. The value chain compresses: supplier → agent → purchase, bypassing ad-supported intermediaries.

Key shifts:

Attention freed — Agents have no human attention bottleneck; can evaluate thousands of options
Collection cost near zero — Agents gather supplier data programmatically
Evaluation at scale — Compare hundreds of options simultaneously
Trust shifts — verify-not-trust replaces brand trust with direct verification

Layer Architecture: Structured vs Negotiated Transactions

Agent commerce splits into two layers based on transaction complexity:

Layer 2 — Structured Transactions

Standardized protocols for agent-to-merchant transactions that follow deterministic rules:

UCP (Universal Commerce Protocol), MCP (Model Context Protocol), agent.json intents
Suitable for: products with clear specs, known prices, repeat purchases, standardized services
Advantage: low latency, deterministic, no LLM inference cost at transaction time
Examples: buying API credits, booking standardized services, purchasing digital goods with fixed pricing

Layer 3 — Negotiated Transactions

Complex scenarios requiring multi-round agent reasoning:

A2A (Agent-to-Agent), A2CN (Agent-to-Commerce-Network), AP2 (Agent Payment Protocol)
Suitable for: bulk purchasing, supply chain negotiations, custom terms, novel purchase situations
Advantage: handles ambiguity, adapts to novel situations, can negotiate terms
Tradeoff: higher latency, LLM inference cost per round, less deterministic

Structured API vs LLM Agent Economics

A key design decision for merchants: should they provide structured APIs or LLM agent interfaces to buyer agents?

Approach	Cost to Merchant	Reliability	Appropriate For
Structured API (UCP/MCP)	Low (no LLM inference)	High (deterministic)	80%+ of standard purchases
LLM Agent Interface	High ($0.01-0.10/query)	Variable	Complex negotiations, custom terms
Hybrid	Medium	High for standard, variable for exceptions	API for standard + agent escalation for exceptions

Implication: Most merchants should default to structured API. LLM agent interfaces are justified only when the deal size exceeds the inference cost — typically B2B procurement, custom manufacturing, or enterprise SaaS negotiations. This mirrors how human e-commerce works: self-serve checkout for standard purchases, sales reps for enterprise deals.

Market Evolution

Traditional Web/API transactions will progressively migrate to Layer 2 as agent adoption grows
Replacement drivers: lower friction, no human attention needed, programmatic comparison at scale
Layer 3 will capture complex B2B share but remain smaller in total transaction count
Both layers are complementary, not competing — structured for volume, negotiated for value

Agent Differentiation Model

Agent capability can be decomposed into four components:

Agent = LLM + Harness + Tool + Memory

Component	Role	Commoditization Risk
LLM	Base reasoning capability	High — rapidly commoditizing across providers
Harness	Orchestration/runtime layer managing state, routing, error handling	Medium — distinct from static workflow and bare runtime
Tool	API access, integrations, domain-specific capabilities	Low — proprietary integrations create moats
Memory	Context retention, learning from past interactions	Low — accumulated user-specific data is unique

Commerce implication: Only differentiated agents create commerce value. Identical agents (same LLM + harness + tools + memory) produce identical decisions — no transaction opportunity. Value arises from:

B2C: Personalization (user preference memory) + unique tool access
B2B: Domain expertise + negotiation capability + proprietary supplier networks

The harness layer is particularly interesting: it's the "operating system" for agents, distinct from both predefined workflows (too rigid) and bare runtimes (too low-level). Companies that own the harness layer (e.g., surf, composio, coinbase-agentkit) capture value regardless of which LLM wins.

Impact on Advertising

Traditional ad models (CPM, CPC, CPS/CPA) were designed for human attention economics. Agents don't see ads, don't click, don't have "impressions."

Emerging patterns:

Agent-readable listings: Structured data agents can evaluate programmatically
Incentive-based discovery: Providers offer better terms for agent-direct purchasing
Verifiable credentials: Replace brand marketing with cryptographically verifiable claims
Procurement-model pricing: Similar to B2B procurement with volume discounts and negotiated terms

B2B vs B2C Dynamics

B2C agent commerce: Likely dominated by Layer 2 structured transactions (standard products, known prices). Personalization and convenience are the value drivers.
B2B agent commerce: More room for Layer 3 negotiated transactions (custom terms, volume discounts, complex requirements). Cost savings and supplier discovery are the value drivers.
Services gap: B2B services procurement online remains relatively underdeveloped vs B2B goods. SaaS subscription procurement via platforms like Vendr and G2 is an early example of agent-mediated B2B procurement.
Cross-border: Both B2C and B2B agent commerce reduce cross-border friction since agents handle currency, compliance, and logistics discovery automatically.

info-intermediary — How information matchmakers evolve in the agent era
verify-not-trust — Trust mechanisms when agents verify instead of trusting
no-app-commerce — Text-to-buy and conversational commerce as precursors
agent-payment-protocols — x402, AP2, MPP payment infrastructure
tempo-mpp — Stripe's agent payment session protocol