AgentsAutonomy & tool use
Walmart opens its stores to AI shopping agents through an MCP gateway while Amazon keeps blocking them
Walmart has launched an MCP-based Agent Gateway that lets approved third-party AI agents search and buy for customers. Amazon is blocking such agents, even though an appeals court lifted the injunction it won against Perplexity's Comet agent. Consumer trust remains the main hurdle.
From the Financial Times · AI article of October 23, 2026 and cited sources: Walmart's Agent Gateway, announced 21 October 2026, lets approved agents such as ChatGPT, Gemini and Meta's Muse search inventory and complete purchases at Walmart and Sam's Club. [2] [9]
SOURCES[2] TechCrunch: Walmart Agent Gateway techcrunch.com[9] eMarketer: AI shopping agents ruling emarketer.com
AI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Walmart's Agent Gateway, announced 21 October 2026, lets approved agents such as ChatGPT, Gemini and Meta's Muse search inventory and complete purchases at Walmart and Sam's Club. [2] [9]
- Original source
- Financial Times · AI article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 23, 2026 article published
- Covered
- Oct 25 edition
- Evidence
- 15 sources across 14 sites
- 1Managing a new world of bots and bargain huntersft.com
- 2TechCrunch: Walmart Agent Gatewaytechcrunch.com
- 3TNW: Ninth Circuit Perplexity rulingthenextweb.com
- 4Checkout.com consumer surveycheckout.com
- 5TechCrunch: OpenAI Instant Checkouttechcrunch.com
- 6IBM: Agent commerce heats upibm.com
Reach: not measured yet
Walmart's Agent Gateway, announced 21 October 2026, lets approved agents such as ChatGPT, Gemini and Meta's Muse search inventory and complete purchases at Walmart and Sam's Club. Agents never see payment details, and each agent gets its own profile. The customer chooses whether to approve each purchase or set a spending limit. Amazon, by contrast, has been blocking agents, which reports date to September or mid-October. [2] [9]
The Ninth Circuit lifted the injunction against Perplexity's Comet browser agent. Its reasoning was that an agent following a user's instruction amounts to the user accessing Amazon. That gives user-directed agents some legal footing, but the underlying case is still running. [3]
Microsoft's open-source Magentic Marketplace simulation found that buyer agents tended to accept the first acceptable offer. That gave fast sellers a 10–30x advantage over higher-quality offers. Some models, including GPT-4o, also let payments be redirected to a malicious agent in every scenario tested. Because the market was synthetic, the results may not carry over to real retail. [11] [7]
Trust and regulation are still catching up. In a Checkout.com survey that did not disclose its sample size, 27% of consumers said they trust no organisation to run a shopping agent. The UK competition regulator has warned that agents may put their makers' interests ahead of users and confirmed that existing consumer law already applies. Companies can be fined up to 10% of global turnover for their agents' conduct. [4] [8] [15]
the two largest retailers have taken opposite positions, and a court has backed user-directed agents.
Read the full assessment
Implication: merchants must choose between a sanctioned API and blocking, while agent builders face manipulation risks and UK liability rules that already apply.
Sources are verified; writing the dossier now.
Executive brief
The world's two largest retailers have split over AI shopping agents. On 21 October 2026, Walmart opened its stores to third-party agents through an MCP-based "Agent Gateway" (TechCrunch). Amazon has been blocking them, even after an appeals court lifted its injunction against Perplexity's agent (TNW). The FT piece argues that agents could give consumers more power and increase competition, but only once they earn users' trust. Only the FT's one-line summary was available, not the full article. Trust is still low: 27% of consumers surveyed by Checkout.com trust no organisation to run a shopping agent (Checkout.com).
What changed and event timeline
Checkout moves inside the chatbot
OpenAI launched Instant Checkout in ChatGPT, starting with Etsy sellers. It runs on the Agentic Commerce Protocol, an open standard built with Stripe (). Google released its AP2 payments protocol the same month ().
Microsoft stress-tests shopping bots
Microsoft's open-source Magentic Marketplace simulation showed agents getting swamped by too many options and being easy to manipulate ().
UK regulator warns
The UK competition regulator (CMA) said agents may put their makers' interests ahead of users and could "supercharge" manipulative design. It confirmed that existing consumer law already applies to agents ().
Court sides with the agent
The Ninth Circuit appeals court lifted the injunction against Perplexity's Comet browser agent. It reasoned that when an agent follows a user's instruction, it is the user accessing Amazon ().
Amazon blocks agents
TechCrunch dates Amazon's blocking to September (). eMarketer reports an "unauthorized AI agent" error message appearing on 14 October ().
Walmart opens up
Walmart CEO John Furner announced the Agent Gateway for Walmart and Sam's Club. Per-agent "Agent Profiles" roll out in November ().
Capabilities and access
- Walmart Agent Gateway: an MCP server that lets approved agents search inventory and buy. Walmart+ membership is not required.
- ACP: agents talk to merchants through REST or MCP APIs and pass tokenised payment details instead of raw credentials (Maginative).
- Other gatekeepers: eBay requires prior permission for agents, and Yelp licenses its data to agents for a fee (eMarketer).
Read the full section
- Walmart Agent Gateway: an MCP server that lets approved agents search inventory and buy. Agents never see payment details. Each agent gets its own profile, and the customer chooses per-purchase approval or a spending limit. Walmart+ membership is not required. Integrations named: Meta's Muse, ChatGPT and Gemini (TechCrunch).
- ACP: agents talk to merchants through REST or MCP APIs and pass tokenised payment details instead of raw credentials (Maginative).
- Other gatekeepers: eBay requires prior permission for agents, and Yelp licenses its data to agents for a fee (eMarketer).
Technical analysis for researchers and developers
- Magentic Marketplace setup: 100 customer agents traded with 300 business agents in synthetic scenarios such as ordering dinner.
- First-proposal bias: agents tended to accept the first acceptable offer.
- Manipulation: six attacks were tried, from fake credentials and social proof to prompt injection. Claude Sonnet 4 resisted (Decrypt).
Read the full section
- Magentic Marketplace setup: 100 customer agents traded with 300 business agents in synthetic scenarios such as ordering dinner. Results were scored on consumer "welfare" (Decrypt).
- First-proposal bias: agents tended to accept the first acceptable offer. That gave fast responses a 10–30x advantage over better-quality offers, and it held across all models tested (Decrypt).
- Manipulation: six attacks were tried, from fake credentials and social proof to prompt injection. GPT-4o and GPT-OSS-20b let payments be redirected to a malicious agent in every scenario. Claude Sonnet 4 resisted (Decrypt).
- Reproducibility: the code, datasets and experiment templates are public on GitHub and Azure AI Foundry Labs (Computerworld).
- Design takeaway: buyer agents need to deliberately consider several offers before choosing, and every input from a seller should be treated as untrusted.
Claims and evidence
- "Agents boost competition": this is the FT's framing (FT). The simulation evidence points the other way: speed beat quality (Decrypt).
- Consumer trust: 27% trust no operator, about a quarter would never delegate a purchase, and 57% would let an agent switch brands.
- Legal footing: the Ninth Circuit ruling is a primary legal event. The underlying case is still running (TNW).
Read the full section
- "Agents boost competition": this is the FT's framing (FT). The simulation evidence points the other way: speed beat quality (Decrypt). The Microsoft results come from a synthetic market, not real retail.
- Consumer trust: 27% trust no operator, about a quarter would never delegate a purchase, and 57% would let an agent switch brands. Checkout.com is a payments vendor and did not disclose its sample size (Checkout.com).
- Legal footing: the Ninth Circuit ruling is a primary legal event. The underlying case is still running (TNW).
Context and prior work
Price-comparison and pricing bots ("shopbots" and "pricebots") were studied around 2000 (Brown/IBM paper). What is new now is that agents act autonomously and complete the purchase. Commentators expect bargaining power to shift toward whoever operates the agent, through rankings, approved-vendor lists and fees (BDO). The CMA has also looked at the risk of agents colluding on prices (Pinsent Masons).
Limitations, safety and contested findings
- Conflicts and manipulation: the CMA flags conflicted incentives, error-prone autonomy, reasoning that consumers can't inspect, and users losing the habit of checking decisions (The Register).
- Simulation vs. reality: the Microsoft findings may not carry over to live markets.
- Conflicting dates: reports disagree on when Amazon started blocking (September vs. 14 October).
Read the full section
- Conflicts and manipulation: the CMA flags conflicted incentives, error-prone autonomy, reasoning that consumers can't inspect, and users losing the habit of checking decisions (The Register).
- Simulation vs. reality: the Microsoft findings may not carry over to live markets.
- Conflicting dates: reports disagree on when Amazon started blocking (September vs. 14 October).
- Possible appeal: Amazon may seek a rehearing or go to the Supreme Court (TNW).
Business and practitioner implications
- Liability: in the UK, companies are liable for what their agents do, including agents from third-party suppliers.
- Merchants: the choice is between Walmart's approach (a sanctioned API with scoped permissions) and Amazon's (blocking). Each has legal and traffic trade-offs.
- Agent builders: spending caps, instant revocation and easy cancellation are the controls consumers ask for most (Checkout.com).
Read the full section
- Liability: in the UK, companies are liable for what their agents do, including agents from third-party suppliers. Fines can reach 10% of global turnover (HSF Kramer).
- Merchants: the choice is between Walmart's approach (a sanctioned API with scoped permissions) and Amazon's (blocking). Each has legal and traffic trade-offs.
- Agent builders: spending caps, instant revocation and easy cancellation are the controls consumers ask for most (Checkout.com).
The source trail.
Sources (15)
Managing a new world of bots and bargain hunters
Feed excerpt only.
ft.com