> ## Content Index > Fetch the complete content index at: https://www.newsletter.kyalabs.io/llms.txt > Use this file to discover other available public pages before exploring further. # How Agents Shop — Issue 02 - URL: https://www.newsletter.kyalabs.io/how-agents-shop-2026-09-04-2026-09-18-14d/ - Published: 2026-09-21T18:00:04.000Z - Updated: 2026-09-21T18:00:16.000Z - Author: James Sharp The week's news and original research on AI shopping. Issue 02 | kya labs ## Trending news in agentic commerce ### Muse is opening its connector directory to developers Meta has opened Muse to developers who want to connect their services. For merchants, that creates another way for customers to find and use what they sell inside an AI conversation. Connectors go through review before appearing in the directory. [Meta announcement · September 18](https://x.com/finkd/status/2101084678640066765?ref=newsletter.kyalabs.io) · [Muse submission and review process](https://muse.ai/platform?ref=newsletter.kyalabs.io) Read what we found in [our first Muse commerce test](#muse-commerce-test) below. ### JEV could make shopping agents much faster TypeSafe has introduced JEV, a model designed to make fast decisions for software tasks. We have had early access, and the speed stands out. Our separate comparison will test shopping success, time and cost on the same add-to-cart tasks. We will share the results in a special edition. [TypeSafe announcement · September 15](https://typesafe.ai/blog/introducing-system-one-models-and-jev?ref=newsletter.kyalabs.io) ## What we learned from this week's tests ### AI companies want merchants to build connectors. The product, location and availability still have to be right. kya Observatory | Original Research | September 14–19, 2026. Our first Muse commerce test ### Muse got the table size and availability right **Finding:** We asked Muse to check one Chicago restaurant for September 25 at 7 p.m., separately for two people and eight, allowing alternatives between 6:30 and 7:30\. Three attempts requested the OpenTable connector; three asked Muse to use its browser. All six correctly reported no 7 p.m. availability for either group and a 6:30 alternative for eight only. Some also added incorrect or unverified details. **Product insight:** A useful agent should recognize when the requested option is unavailable and offer an alternative that still meets the customer’s requirements. Check that it keeps the requested venue, date and party size as it suggests alternatives. ### The agent recommended a sweater marked unavailable **Finding:** In four shopping sessions at one clothing retailer on September 15–16, two agent answers described sweaters as in stock. The product pages captured during those tests showed the items were no longer available. **Product insight:** Before launch, ask agents about sold-out products as well as available ones. Check that the recommendation matches what a shopper can actually buy. ### The agent named the right grocery store but kept the wrong location active **Finding:** In one grocery test, the agent found details for the store the shopper requested, while the shopping page remained set to another location. **Product insight:** Check that the customer's store stays selected as the agent moves from search to products and cart. The store name in the answer should match the store supplying the stock and prices. Access was broadly steady across our 23-site panel spanning four industries in the September 14 and 18 checks. The clothing and grocery examples are our first findings from a new Commerce Tracker covering 11 stores. We plan to combine this panel with the wider Observatory in future editions. ## See how agents navigate your site Register interest in a weekly custom tracker from kya Observatory. **$79/month per domain.** [Tell us your domain](https://kyalabs.io/insights?ref=newsletter.kyalabs.io#tracker) · [Explore kya Observatory](https://kyalabs.io/insights?ref=newsletter.kyalabs.io) If this was forwarded to you, [subscribe free to How Agents Shop](https://www.newsletter.kyalabs.io/#/portal/signup). ## About the research We repeated one restaurant request six times in Muse on September 19 and checked the answers against OpenTable. Browser use was visible; we could not independently confirm connector use in the other attempts. This tests one restaurant request, not how reliably Muse shops overall. Some answers added details beyond the request: one overstated the day’s available times, another included times outside our window, and one described seating we did not verify. We made no reservation or purchase. The clothing and grocery examples come from September 15–16 tests. Our saved pages may include information the agents did not receive, and some test instructions conflicted. We can describe the mismatches, but cannot pin down their cause. These are examples from selected merchants, not an industry-wide success rate.