Explainer Collection

Agentic Commerce

Yu, Xiao, Zhao, Luo & Zeng (2026)

AI shopping agents have a memory problem - and Alibaba just solved it

A new benchmark spanning 1.2 million real products reveals that even state-of-the-art models struggle to remember what you actually want. A lightweight memory-augmented agent trained end-to-end beats them all.

1.2M: Real-world products in the new benchmark dataset
<70%: Success rate for top-tier models including GPT-5

Stripe (2026)

AI agents are your new customers - and your website probably can't serve them yet

Stripe's technical field guide for agentic commerce explains the concrete infrastructure changes businesses need to make to be discovered, parsed, and transacted with by AI shopping agents - from robots.txt configuration to ACP-compliant checkout endpoints.

ACP: Agentic Commerce Protocol - the open standard Stripe and OpenAI co-developed for agent-native commerce
3-layer: Discovery → Access → Transaction: the three infrastructure layers agents need to complete a purchase

Stripe (2026)

Stripe built agentic commerce from scratch - here are the 10 things that surprised them

Since co-developing the Agentic Commerce Protocol with OpenAI in September 2025, Stripe has shipped four protocol releases and onboarded leading brands. Their honest retrospective covers catalog fragmentation, real-time data demands, scoped tokens, and why sellers can't bet on a single protocol.

6: Catalog formats brands had to maintain across different AI agents - before Stripe's unified suite
4: ACP protocol releases since September 2025: payment handlers, scoped tokens, discounts, buyer auth, MCP transport

Navakoti & Navakoti (2026)

Stop documenting everything upfront - let agent failure tell you what knowledge to build

Demand-Driven Context (DDC) is a TDD-inspired methodology for building enterprise knowledge bases: give an AI agent a real problem, let it fail, then curate only the minimum knowledge needed for it to succeed. Nine cycles produced 46 reusable entities for a retail SRE agent.

46: Knowledge entities produced in 9 DDC cycles for one SRE domain role
20 - 30: Cycles until knowledge base convergence for a given domain role

Kim & Kim, Arizona State University (2026) · WWW '26

OAuth gives AI agents too much power - this paper designs a website that gives them exactly the right amount

Current e-commerce websites have no standard way to delegate transactional authority to AI agents with fine-grained control. Kim & Kim extend an open-source authorization toolkit with cryptographic delegation tokens and a three-tier trust system - and open-source the whole implementation.

3: Trust levels for AI agents: High, Medium, and Low - each with distinct validity windows
9-step: Cryptographic delegation workflow between user, agent, Auth, and website

Gal Bakal (2026) · Preprint

The bottleneck isn't your AI model - it's that your institutional knowledge isn't built for agents to consume

Knowledge Activation introduces Atomic Knowledge Units (AKUs): action-ready specifications that encode what to do, which tools to use, and where to go next - forming a traversable knowledge graph that compresses onboarding and eliminates the correction cascades that burden senior engineers.

AKU: Atomic Knowledge Units - the proposed institutional knowledge primitive for the agentic era
Graph: AKUs form a composable knowledge graph that agents traverse at runtime - not a flat document store