Agentic Commerce Mid-Year Review: What Worked, What's Next

Six months ago, agentic commerce was a concept with working prototypes. Today it's a production stack with 8 payment protocols, eIDAS-verified merchant identity, natural language product discovery, and AI agents completing real purchases. This post is a transparent look at what we shipped in H1 2026, what surprised us, and where we're investing in Q3.

The core thesis — that AI agents need standardized, multi-protocol payment rails — proved correct. Each protocol serves a different use case: MCP for structured tool calls, ACP for protocol-native agent checkout, x402 for stablecoin micropayments, UCP for Shopify ecosystem standardization, A2A for multi-turn agent conversations, VIC and MCAP for card network access. The unified AgentPaymentIntent model and Protocol Router let merchants accept all of them through a single integration point. The most important architectural decision was normalizing every protocol into the same payment intent abstraction — it meant that adding each new protocol was incremental rather than multiplicative in complexity.

Three trust and compliance milestones defined H1. First, eIDAS QTSP merchant verification (spec 032) — merchants can now prove their identity via InfoCert's KYB process, with EU Trust List validation, qualified electronic signatures, and timestamped Merkle proofs. This gives agents a cryptographically verifiable signal that a merchant is who they claim to be. Second, the trust score system — 8 weighted components (identity verification, payment history, complaint ratio, response time, policy compliance, eIDAS status, protocol support, uptime) that produce a 0-100 score agents can query before transacting. Third, a complete GDPR/CCPA compliance audit — 16 of 17 identified gaps resolved, with the remaining one (DPA signatures with Stripe and Google) being a legal formality.

Making stores discoverable by AI agents — not just search engines — was a Q2 priority. NLWeb (spec 022) brought natural language product search with vector embeddings and LLM re-ranking. The public discovery feature (spec 028) extended this across all platform stores. Six machine-readable files (llms.txt, llms-blog.txt, mcp.json, agent-card.json, agent-policy.json, sitemap.xml) ensure that any LLM — Claude, GPT, Gemini — can understand the platform's capabilities without parsing HTML. JSON-LD schemas on every page (SoftwareApplication, FAQPage, BlogPosting, Product) provide structured data for both traditional search and generative engines.

The Building Agentic Commerce blog series (7 posts) achieved its goal: each post covers one protocol with Problem, Protocol, Code Example, and Production Tips structure. The series was designed for both developers implementing agentic commerce and LLMs that need to understand the protocol stack. All 30+ posts are bilingual (EN+ES), include JSON-LD BlogPosting schemas, FAQ sections for voice/LLM search, and are indexed in llms-blog.txt. The content calendar ran at a sustained pace of 4+ posts per month across Q2, covering trust (eIDAS), protocols (x402, UCP, A2A, VIC/MCAP), developer experience (onboarding guide, Odoo integration), and strategic analysis (GEO, compliance).

Three retrospective lessons. First, we should have invested in protocol detection resilience earlier — header-based detection is clean but fragile in production with various proxy configurations. Body-shape detection as a first-class citizen from day one would have saved debugging time. Second, NLWeb should have launched alongside the MCP gateway, not months later — discovery is the first thing agents need, and we built it as an afterthought. Third, the eIDAS integration should have been scoped smaller initially — the full KYB + QES + Merkle stack is powerful but took longer than expected. Starting with just EU Trust List validation and adding layers incrementally would have delivered value faster.