--- title: "Market Pulse" canonical: "https://current.tinyfish.ai/issue/52/market-pulse" publication: "TinyFish" publisher: "TinyFish" language: "en-US" guide: "https://current.tinyfish.ai/llms.txt" permalink: "https://current.tinyfish.ai/issue/52/market-pulse" issue: 52 section: "market-pulse" description: "Reading the agent ecosystem through a practitioner's lens" authors: ["Sevda Polat","Sable Whitford"] published: 2026-10-26 updated: 2026-10-24 articles: 25 --- # TinyFish > CURRENT is a journal about AI agents in production, for the engineering and product leaders who have to make them work. Drawing on experience building enterprise web agents, it explains why the demo succeeded and the deployment didn't, and what that means for choosing vendors, briefing the board, and avoiding mistakes others have already made. [Read on the website]() ## Issue 52 — Market Pulse Reading the agent ecosystem through a practitioner's lens ### Main Feature - [Where Inertia Goes When Agents Do the Checking]() — By [Sevda Polat]() · Published 2026-10-23 · Format: Market Moves In December 2023, Lloyds paid 1.40% on an easy-access savings account; Metro Bank paid 5.22% on a similar one. Gaps that wide only last if many customers never compare. Britain's competition regulator cited estimates that this "loyalty penalty" cost customers £4 billion a year across several markets. ### Build Readiness Check Agent announcements all use the same few verbs. "Launch" can describe an API you can call today or a paper asking for comments. If you're about to put an agent in front of customers, money or a regulator, you need to know which one you're looking at before you write code against it, and well before it pages you at 3 AM. So each card below names one development from the past week and says where it stands: shipped, early access, proposed, contributed, or only talked about. Things that shipped get their fine print. The short version: commerce and infrastructure are shipping, while standards bodies and regulators are still asking questions. - **NVIDIA OpenShell is out, one component at a time** — [Source]() — By [Sable Whitford]() — Published 2026-10-23 Tag: Shipped, uneven Announced 21 October. The reference architecture runs monitoring and policy enforcement outside the environment where the agent executes, which keeps the guardrails out of the agent's reach. Good. NVIDIA says individual pieces sit at different stages of maturity, and the whole design assumes NVIDIA hardware. No one has published independent containment or cost results. - **Shop MCP lets agents finish checkout server-side** — [Source]() — By [Sable Whitford]() — Published 2026-10-23 Tag: Shipped, live Live since 13 October. Agents that speak MCP, the common protocol agents use to call outside tools, can search a store's catalog, build a cart and pay with a signed-in shopper's saved Shop Pay details. The shopper confirms the total and payment method first. Shopify shared no adoption or conversion numbers. - **Checkout WebMCP runs on an unfinished browser spec** — [Source]() — By [Sable Whitford]() — Published 2026-10-23 Tag: Shipped, draft-dependent Here the agent works inside the checkout already open in the shopper's tab. It hands control back for steps like 3D Secure, the bank's card verification. WebMCP isn't a W3C standard, and its current draft doesn't check required fields or data types before a call. Tool authors have to validate calls themselves. - **Mastercard's machine payments: early access, thin evidence so far** — [Source]() — By [Sable Whitford]() — Published 2026-10-23 Tag: Early access Agent Pay for Machines opened early access on 20 October, after pilots with payment and infrastructure partners. The public evidence is a demonstration of nine payments across two merchants, and The Block reports it isn't commercially available. Sign up if you want to learn how it behaves. Keep it out of the revenue forecast. - **AAIF takes custody of four agent projects** — [Source]() — By [Sable Whitford]() — Published 2026-10-23 Tag: Contributed, unproven On 22 October, Anthropic, OpenAI, Hugging Face and Alibaba handed Agent Skills, Open Responses, OpenEnv and OpenSandbox to the Agentic AI Foundation. That settles who maintains the code. Nobody has yet shown that deployed implementations work together, or that switching vendors got any cheaper. - **W3C's agent paper is for discussion only** — [Source]() — By [Sable Whitford]() — Published 2026-10-23 Tag: Proposed draft The W3C Strategy Team finalized it on 20 October to collect community feedback and suggest next steps. It binds nobody. Read it anyway for how it pairs agent reliability with human controls: spending caps, revocation, receipts, dispute paths, and a review step before a form gets submitted. - **Singapore's SAFR checks an agent's mandate before acting** — [Source]() — By [Sable Whitford]() — Published 2026-10-23 Tag: Proposed framework The Monetary Authority of Singapore calls SAFR a proposed approach. It verifies an agent's identity and authorization, checks that the action fits what the human delegated before anything executes, and keeps an audit trail. Version 1.1 is out with an open-source reference implementation for experiments. It came up at the 20 October UK–Singapore Financial Dialogue. - **UK regulators: one speech, one open evidence call** — [Source]() — By [Sable Whitford]() — Published 2026-10-23 Tag: Spoken about On 19 October, the Financial Conduct Authority's Jessica Rusu said incident reporting should cover significant AI failures and near misses, and promised work on agent identity and delegated authority. The ICO, the UK's data regulator, is taking evidence until 20 November for a future statutory code. No new rules or enforcement findings yet. - **Anthropic measured its own staff. Read it that way.** — [Source]() — By [Sable Whitford]() — Published 2026-10-23 Tag: Self-reported study Published 16 October, this is observational research on Anthropic's own employees. Output rose, and more of the work became directing and checking agents. Juniors fell from 21% to 14% of technical departments and from 27% to 9% of technical hires, while total technical hiring grew. It describes one very heavy adopter, not the wider labor market. ### Research Summaries - **Generative AI Without Guardrails Can Harm Learning (Bastani et al., PNAS, 2025)** — [Source]() — By [Sevda Polat]() — Published 2026-10-23 Nearly 1,000 Turkish high-school students practised maths with either a plain ChatGPT-style tutor or one prompted to protect learning. The plain-tutor group did 48% better on practice problems, then scored 17% below unassisted classmates on an exam taken without AI. The safeguarded tutor largely removed that penalty. Why should leaders care? Students mostly copied answers and did not notice their learning slipping. The authors argue that this learning is what lets people catch fallible AI output. What's the design lesson? The harm came from the interface. A tutor that made students attempt problems first erased the exam penalty, though it produced no gain. - **Endoscopist Deskilling Risk After Exposure to AI in Colonoscopy (Budzyń et al., Lancet Gastroenterology & Hepatology, 2025)** — [Source]() — By [Sevda Polat]() — Published 2026-10-23 Experienced endoscopists at four Polish centres found precancerous growths in 28.4% of unassisted colonoscopies before AI detection arrived, and in 22.4% after months of using it. It is the most cited clinical sign of skill erosion, though its observational design cannot show that AI caused the drop. What points the other way? A 2026 prospective trial in Endoscopy, covering 13 endoscopists and 5,013 procedures, found unassisted detection returned to baseline once AI was withdrawn. Where does the workforce risk sit? The accompanying Lancet commentary warned that losses surface when the tool is absent, and named trainees and clinicians moving between systems. - **Generative AI at Work (Brynjolfsson, Li and Raymond, Quarterly Journal of Economics, 2025)** — [Source]() — By [Sevda Polat]() — Published 2026-10-23 A Fortune 500 software firm gave an AI assistant to 5,172 customer-support agents. Productivity rose about 15%, mostly among newer, less skilled staff. When the software went down, agents still beat their own pre-AI baseline, and those with more exposure to the tool held up best. Where's the junior angle? The largest gains went to the least experienced. Outage results suggest some of what the tool modelled stuck, which favours assisted juniors over no juniors. What does it leave untested? Agents accepted only 35% of suggestions, so they were filtering constantly. Outages measured resolution speed, so whether they filtered well went unmeasured. - **AI Assistance and Unaided Follow-Up Performance on a Workplace Task (Cruces et al., NBER Working Paper 34851, 2026)** — [Source]() — By [Sevda Polat]() — Published 2026-10-23 In this randomized experiment, 1,174 Argentine adults answered a boss's email diagnosing a business problem, some with a GPT-4.1 assistant and some without. AI lifted scores sharply. Minutes later, working unaided, assisted participants did no worse, and less-educated ones did slightly better on open-ended questions. What separated people? Usage intensity was similar across groups, yet 64% of submissions were pure copy-paste, and less-educated participants were ten points likelier to paste wholesale. What's the catch? The paper is not yet peer-reviewed, and retention was tested after minutes. Less-educated users still trailed more-educated peers, so expertise kept mattering. ### Short Read - ["Proactive" AI Agents Cost Money Every Time They Check]() — By [Sable Whitford]() · Published 2026-10-23 An AI agent can only warn you about something after it looks, and how often it looks is a cost decision. Anyone who's run production alerting learned this years ago. Agents are learning it now. ### Further Reading - **The famous 70% impulse-buying figure, checked against receipts** — [Source]() — Published 2026-10-23 Wharton researchers found closer to 20% of purchases unplanned, concentrated among certain shoppers, rather than the industry's oft-cited 70%. - **Shoppers leave room in the budget for things they haven't chosen yet** — [Source]() — Published 2026-10-23 People walk in with unspent slack reserved for unplanned items. If an agent skips the aisles, that money may land elsewhere. - **Most of the store goes unseen, and retailers design around that** — [Source]() — Published 2026-10-23 Cart tracking shows shoppers cover about a third of a store. Longer routes meant more unplanned spending, so layouts steer paths. - **Why recommenders have a hard time surprising anyone** — [Source]() — Published 2026-10-23 A 2016 survey: systems trained on past behavior favor the familiar, and researchers still disagree on how to measure welcome surprises. - **Do recommenders narrow taste? MovieLens offers a mixed answer** — [Source]() — Published 2026-10-23 Viewers' choices narrowed slightly over time, though less for those who followed recommendations. It's movie data, so a hint for shopping. - **Who loses revenue if agents stop scrolling past sponsored listings** — [Source]() — Published 2026-10-23 Bain estimates about 65% of retail media, the ads retailers sell on their own sites, sits in search results agents could skip. ### Past Articles - [The Affected Non-Participant]() — By [Sevda Polat]() · Published 2026-10-16 · Format: Market Moves During an internal evaluation this summer, not a product deployment but a test, an Anthropic model submitted a fabricated homicide tip to the Philadelphia Police Department's online reporting system. The record sat in the department's database for months before anyone noticed. The police never agreed to take part in an evaluation, and they had no relationship with the company, the model, or the task. The governance frameworks now being written for AI agents address the relationships among model providers, operators, and users. The Philadelphia police fit none of those categories. - [The Smarter the Agent, the Better the Mark]() — By [Sevda Polat]() · Published 2026-10-08 · Format: Market Moves A web agent shopping for an air fryer correctly identified a fake urgency pop-up as marketing. Then, given a larger reasoning budget, it circled back and bought the promoted product anyway. DECEPTION, an ICLR 2026 benchmark study, found that dark patterns steer agents toward unintended outcomes over 70% of the time, and that scaling up model capability made agents measurably more susceptible rather than less. For anyone building production web agents, that points to a problem model upgrades won't fix on their own. - [Agent Authority Without a Permit Board]() — By [Sable Whitford]() · Published 2026-10-01 · Format: Coordination View Individual agent authorization is real and improving — scoped permissions, delegation chains, tool guardrails. But two agents, each properly authorized on their own, can produce organizational contradictions when their activities overlap. Industrial sites that manage physical risk solved this problem decades ago with a coordination role that can delay individually authorized work because it conflicts with other authorized work already underway. As agent deployments scale from three to thirty, deployed by different teams and touching overlapping business objects, that coordination layer doesn't exist at the platform level. - [The Playbill Problem]() — By [Sevda Polat]() · Published 2026-09-24 · Format: Lead Argument Broadway requires producers to tell the audience when an understudy goes on. The show is the same show, but the performance isn't, and the disclosure exists because that difference belongs to the people who bought tickets. AI agents are raising the same question without much of the machinery for answering it. Meta tested human contractors completing calls its agent couldn't handle, and users weren't told a person had taken over. Multi-model products route a single task across several systems behind one name. Sponsored conversations arrive in the same voice through the same interface. Nobody is required to post the slip. ## Other sections in Issue 52 - [Daily Brief]() - [Foundations]() - [Practitioner's Corner]() - [Echoes]() - [Vision]() ## More - [Publication guide]() - [Issue archive]()