Sunday, August 2, 2026

— A weekly publication —

The Agentic Commerce Report

A weekly read of everything that moved in agentic commerce — protocols, payment rails, retailer pilots, regulation. Summarised, sourced, and stitched to what came before.

Google Builds Computer Use into Gemini 3.5 Flash as Agent Pay Reaches Ukraine

Issue 29June 22–28, 2026Synthesised from 4 sources

Edited by Reviewed against primary sources

Google embedded computer use as a native tool in Gemini 3.5 Flash on June 24, making a capability previously limited to a separate preview model available in its fastest, lowest-cost production model 3. Gemini 3.5 Flash, priced at $1.50 input and $9 output per million tokens, runs at approximately 30% of GPT-5.5’s cost and scores 78.4 on OSWorld-VerifiedA benchmark measuring autonomous desktop and browser control by AI agents, including multi-step task completion and checkout flows. (a benchmark measuring autonomous desktop and browser control by AI agents) — within 0.3 points of GPT-5.5 and 0.4 above Claude Opus 4.7 3. Computer use lets agents navigate browser and desktop interfaces to complete multi-step actions including cart interactions and checkout flows without requiring structured API access from merchants. Google extended the context window for computer-use sessions to one million tokens, up from the prior 128K cap. Four days later, Google’s Interactions APIGoogle's unified interface for Gemini models and agents, replacing generateContent as the primary endpoint for new agent development. (the unified interface for Gemini models and agents) reached general availability 4. It consolidates server-side session state, observable execution steps, and background-task execution into a single endpoint, replacing generateContent as the primary development interface.

Both announcements extend an agent model stack that has moved quickly since May 2026. Claude Opus 4.8 set a new Online-Mind2WebA benchmark measuring autonomous web navigation capability, including multi-step checkout flows, used to evaluate browser-agent model performance. (a benchmark for autonomous web navigation including checkout-flow completion) high of 84% that month (2026-w22); Gemini 3.5 Flash’s 78.4 OSWorld-Verified score adds a cost-optimized alternative at the same tier. The Interactions API 4 becomes the only path to Managed AgentsGemini agents that run in dedicated Linux sandboxes and support background execution of long-running tasks via the Interactions API. (Gemini agents in dedicated Linux sandboxes with background-mode execution) and to any future frontier capabilities; Google explicitly designates generateContent as legacy. Server-side state in the Interactions API reduces token overhead across multi-turn agent sessions. Stripe’s Radar bot-score layer (2026-w22) and Visa’s DCAP authorization program (2026-w23) provide the fraud and settlement infrastructure downstream of these model-and-API upgrades.

Mastercard and PrivatBank completed Ukraine’s first agentic payment on June 23 1. PrivatBank, Ukraine’s largest state-owned bank with approximately 22 million active customers, confirmed the transaction through Mastercard’s Agent Pay framework, which attaches a tamper-resistant consent record to each agent-initiated transaction for issuer visibility. Ukraine marks the seventh confirmed live market for Agent Pay. Three of those confirmed markets went live in June 2026 alone (2026-w23, 2026-w24, 2026-w26). The Ukraine deployment runs outside the EU regulatory perimeter and PSD2 strong customer authentication requirements, adding a distinct compliance environment to the confirmed live set. Also on June 23, Anthropic launched Claude Tag 2 — a Slack-resident async agent powered by Opus 4.8 that joins a channel as a single shared team member. Claude Tag connects to data sources via Model Context Protocol (MCP, the open standard for connecting agents to external tools and data) servers that administrators grant per-channel access to. Anthropic retires its existing Claude in Slack integration on August 3, 2026.

Google and Anthropic each consolidated a model-and-deployment stack within the same seven days without explicit coordination. Google ships Gemini 3.5 Flash for computer-use execution and the Interactions API for state management and agent orchestration. Anthropic ships Opus 4.8 for browser-agent tasks and Claude Tag for persistent enterprise deployment, with MCP governing tool access. Adyen Agentic, announced June 16, functions as an integration layer for enterprise merchants. It bridges Universal Commerce Protocol (UCPUniversal Commerce Protocol — Google's open standard for agent-initiated product discovery, cart interactions, and checkout across retailers., Google’s open standard for agent-initiated product discovery and checkout), Agentic Commerce Protocol (ACPAgentic Commerce Protocol — OpenAI and Stripe's agent checkout framework for AI-initiated purchases across merchant platforms., OpenAI and Stripe’s agent checkout framework), and AP2Google's Agent Payments Protocol — a draft open standard for delegated, agent-initiated payment flows. (Google’s Agent Payments Protocol for delegated payment flows). No jurisdiction has enacted regulation specific to agentic commerce as of this issue. The EU AI Act enters full application on August 2, 2026, imposing risk-based requirements on AI systems; the Act predates autonomous agent purchasing as a defined transaction category.

Events this issue

7 events
Standards
launch

Interactions API: our primary interface for Gemini models and agents

Google's Interactions API for Gemini models and agents reached general availability on June 28, replacing generateContent as the primary interface.

The Interactions API entered public beta in December 2025 and adds three capabilities over the legacy generateContent endpoint: server-side conversation state via previous_interaction_id, observable execution steps for debugging and UI rendering, and background execution for long-running tasks. Server-side state management increases context cache hit rates across multi-turn agent sessions, reducing per-token costs for commerce agents that run extended discovery and comparison workflows. The API is the sole pathway for Managed Agents — agents with dedicated Linux sandboxes — and for background-mode tasks expected to run beyond a single API call. Google explicitly classifies generateContent as legacy; new frontier capabilities, including future agent tool integrations, will land exclusively on the Interactions API. The GA announcement follows computer use integration into Gemini 3.5 Flash earlier in the same week (2026-w26), completing both model capability and API interface upgrades within seven days.

  1. Google Blog
AEO
launch

Google AI Mode expands store product availability checker to let agents call physical stores on users' behalf

Google extends its AI-powered store-calling feature — using Gemini models and Duplex technology — to AI Mode, enabling agents to verify local in-store product availability and deals.

Google's in-store availability checker in AI Mode marks the extension of agentic commerce from e-commerce to omnichannel retail: the agent closes the loop on a product search by autonomously calling physical stores to confirm stock before the user commits to visiting. The underlying technology — Gemini models layered on Duplex's conversational telephony stack — represents a decade of investment in voice-mediated commerce converging at AI Mode alongside Google's agentic booking rollout (2026-w26-travel-google-ai-mode-ota-booking-partners). For brick-and-mortar retailers, the feature surfaces inventory to AI-mediated demand without requiring API integration with Google Shopping or the Universal Commerce Protocol (2026-w02-standards-google-ucp-launch): any store with a phone line becomes accessible to the agent. This creates agentic commerce reach for the long tail of retailers with no structured catalogue connectivity, fundamentally changing the value of local store presence in an agentic search landscape.

  1. Google Blog
Pilots
pilot

Papa Johns tests Gemini AI agents for agentic ordering and loyalty programme management alongside Kroger

Papa Johns joins Kroger in deploying Google's Gemini Enterprise for CX agents, testing AI-driven ordering and loyalty redemption in a quick-service restaurant context.

Papa Johns' Gemini deployment extends Google's agentic commerce stack into quick-service restaurants, a high-volume, high-frequency category with different constraints than grocery planning or luxury gifting. Together with Kroger (2026-w02-grocery-kroger-google-gemini-shopping-agent), the simultaneous deployment demonstrates that Google is building a cross-sector agentic retail portfolio on Gemini Enterprise for CX, using the Universal Commerce Protocol (2026-w02-standards-google-ucp-launch) as the commerce layer beneath each deployment. QSR ordering is structurally simpler than grocery — finite menus, no perishables, faster delivery cycles — making it a viable test environment for agentic loyalty integration: an agent that can autonomously apply points, apply a promotional offer, customise to dietary preferences, and complete payment in a single conversational turn. If successful in QSR, the pattern is directly portable to grocery click-and-collect and CPG subscription replenishment models at Kroger and beyond.

  1. PYMNTS
Standards
launch

Introducing computer use in Gemini 3.5 Flash

Google embedded computer use natively into Gemini 3.5 Flash on June 24, scoring 78.4 on OSWorld-Verified at $1.50 per million input tokens.

Gemini 3.5 Flash is Google's fastest, most economical production model, priced at approximately 30% of GPT-5.5's cost. Prior computer-use capability was limited to a separate Gemini 2.5 Pro preview with a 128K token context cap; the June 24 update integrates it into the main production model with a one-million-token context window. Computer use lets agents navigate browser and desktop environments to complete multi-step actions — including cart interactions and checkout flows — without requiring structured API access from merchants. Two configurable enterprise safeguards are available: explicit user confirmation before irreversible actions, and automatic session termination on detected prompt injection. The 78.4 OSWorld-Verified score places Gemini 3.5 Flash within 0.3 points of GPT-5.5 (78.7) and 0.4 above Claude Opus 4.7 (78.0). The launch extends the agentic model infrastructure layer built on Claude Opus 4.8's Online-Mind2Web benchmark results (2026-w22).

  1. Google Blog
Pilots
pilot

Mastercard and PrivatBank complete Ukraine's first payment executed by an AI agent

Mastercard and PrivatBank completed Ukraine's first Agent Pay transaction on June 23, adding a seventh live market.

Ukraine extends Mastercard's live Agent Pay map, which previously confirmed deployments in Australia (January 2026), India and New Zealand (February), South Korea and Latin America (March), and the Netherlands via ING and Worldline (2026-w23). Unlike the Netherlands transaction, the Ukraine deployment runs outside the EU regulatory perimeter and PSD2 strong customer authentication requirements, adding a distinct regulatory environment to the confirmed set. PrivatBank is Ukraine's largest state-owned bank, serving approximately 22 million active customers; it is the first public-sector bank to confirm a live Agent Pay transaction. Mastercard's framework attaches a tamper-resistant consent record to each transaction for issuer visibility across the authorization chain. No specific merchant category, transaction value, or volume was disclosed. Three confirmed Agent Pay deployments occurred in June 2026 alone (2026-w23, 2026-w24, 2026-w26).

  1. Mastercard Newsroom
Standards
launch

Introducing Claude Tag

Anthropic launched Claude Tag on June 23, a Slack-native async agent powered by Opus 4.8 with MCP-connected tool access.

Claude Tag joins a Slack channel as a single shared agent instance: all participants delegate tasks to the same Claude, which posts threaded milestone updates and retains channel context across sessions. The product replaces Anthropic's existing Claude in Slack integration, retired on August 3, 2026, with a 30-day migration window from the June 23 launch. Claude Tag connects to GitHub repositories, internal documents, and SaaS applications via Model Context Protocol servers administrators grant per-channel access to. An optional ambient mode allows Claude to post unprompted when it identifies relevant cross-channel information. Anthropic reports its own product team uses an internal Claude Tag build to generate 65% of their code, including the code that built Claude Tag itself. Available in research preview for Enterprise and Team customers. Fable 5 and Mythos 5 remain suspended under the June 12 US government directive (2026-w24).

  1. Anthropic News
Pilots
launch

Google AI Mode announces agentic hotel and flight booking with Booking.com, Expedia, Marriott, IHG, and Wyndham

Google confirms six major OTA and hotel brand partners for agentic in-AI-Mode booking, with full launch pending as booking flow and payment integration are finalised.

Google's OTA partner announcement represents the largest potential agentic travel distribution channel in history: Google Search processes more travel queries than any platform, and AI Mode converts those queries into agent-executable booking intents rather than link recommendations. The six launch partners — Booking.com, Choice Hotels, Expedia, IHG, Marriott, and Wyndham — collectively control the majority of global branded hotel inventory. The announcement pre-launches the infrastructure before the payment and booking rails are fully live, signalling Google's intent to establish its agentic travel standard before Sabre's stack (2026-w19-travel-mindtrip-flights-launch) or individual OTA agents reach comparable distribution. If Google enables checkout without leaving AI Mode, it disintermediates OTA booking fees and fundamentally reshapes how travel inventory reaches consumers — the deepest structural threat to Booking.com's $23B annual revenue in the company's history.

  1. Search Engine Journal