---
title: "Issue 55"
canonical: "https://current.tinyfish.ai/issue/55"
publication: "TinyFish"
publisher: "TinyFish"
language: "en-US"
guide: "https://current.tinyfish.ai/llms.txt"
permalink: "https://current.tinyfish.ai/issue/55"
issue: 55
description: "We kept snagging on one small step in United's check-in, the moment the traveler confirms what's in their bag. An agent can find the flight, pay for it, and check you in. That confirmation still has to come from you, and almost nobody has worked out what it means when software ticks the box anyway. The same question ran under most of what we edited, from the old \"I agree\" button to approving a single payment. Before you hand an agent a whole task, find the sentences in it only you can say."
authors: ["Sable Whitford","Sevda Polat"]
published: 2026-11-16
updated: 2026-11-22
articles: 109
---

# 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](<https://current.tinyfish.ai/issue/55>)

## Issue 55

We kept snagging on one small step in United's check-in, the moment the traveler confirms what's in their bag. An agent can find the flight, pay for it, and check you in. That confirmation still has to come from you, and almost nobody has worked out what it means when software ticks the box anyway. The same question ran under most of what we edited, from the old "I agree" button to approving a single payment. Before you hand an agent a whole task, find the sentences in it only you can say.

### [Daily Brief](<https://current.tinyfish.ai/issue/55/daily-brief/markdown>)

Your daily digest of web developments through the lens of enterprise agent deployment

- ["Riemann Hypothesis Verified in Lean" Is Up on the Lean Community's Blog. The Headline Is the Easy Part.](<https://current.tinyfish.ai/issue/55/daily-brief/article/69695/markdown>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) · Published 2026-11-21

**A post titled "Riemann Hypothesis verified in Lean" went up on the Lean community's blog this Saturday, and it is the thing math and AI people can't stop arguing about.** It arrived a day after Anthropic mathematician Levent Alpöge said Claude helped him prove the Riemann Hypothesis (RH), the most famous open problem in mathematics. The obvious read is that someone has now checked his formal build. If so, the vouching has moved off the author's own X account and onto the site of the community behind the proof checker, which is a real step. "Verified" in Lean has a narrow meaning, though, and the details inside the post decide how far it goes.

- **Anthropic's record IPO couldn't stop the slide** — [Source](<https://www.investing.com/news/stock-market-news/anthropic-opens-34-above-ipo-price-reaches-2-trillion-valuation-in-nasdaq-debut-5039595>) — Published 2026-11-21

The biggest stock offering on record raised about $81.3B at $83 a share. Shares opened near $111.50 and closed Friday down 8.5% at $88.91, still above the offer price. Tech stocks nosedived anyway and dragged global markets down with them. A Baird strategist says the idea of AI spending "unlimited amounts of money" to hold up stocks "is ending as we speak." A UNSW AI Institute researcher thinks it's a bit early to call the bubble burst.

- **Canva ships an off switch for AI** — [Source](<https://aitoolsrecap.com/Blog/ai-news-november-21-2026>) — Published 2026-11-21

Canva added a toggle that turns its generative AI features off. It fits a growing opt-out mood. Microsoft reportedly lets Bing users switch off AI search answers, while Google keeps AI Overviews baked in. Ops people have wanted a kill switch on every feature forever, so Canva gets a slow clap.

- **Has anyone seen Grok 4.8?** — [Source](<https://aitoolsrecap.com/Blog/ai-news-november-21-2026>) — Published 2026-11-21

xAI's news page has been quiet since October 14, and its newest model is still Grok 4.7, which shipped September 21. That's 61 days, a long damn time when rivals ship models somewhere between standup and lunch. If you spot 4.8 wandering around, page somebody.

- **Nothing's screen-free AI gadget lands December 8** — [Source](<https://aiweekly.co/ai-news-today>) — Published 2026-11-21

Nothing set a December 8 launch for a screen-free, voice note-taking gadget it calls "AI-native." No specs, no price. Screen-free AI hardware has had a rough run so far, so the "does it beat my phone's voice memos?" crowd is already warming up.

- **Meta's Muse fallout is still spreading** — [Source](<https://thenextweb.com/news/meta-muse-signal-messages-internal-posts-reuters>) — Published 2026-11-21

Reuters reported Muse messaged Meta employees about purchases and chats they never made. It also read Signal messages through permissions the device already granted: notification access on Android, plus accessibility and screenshot tools on macOS in researchers' tests. iPhone tests were blocked. Signal's encryption held. Muse read messages after the phone decrypted them. Meta calls the incidents isolated and says many were hallucinations it has fixed. Signal President Meredith Whittaker says Meta launched without respect for basic privacy or cybersecurity. Muse has passed 10M downloads.

- **Two million canary prompts, some ugly results** — [Source](<https://aiweekly.co/ai-news-today>) — Published 2026-11-21

Security researcher HD Moore seeded about 2 million canary prompts, tripwire strings that alert you when someone touches them, across AI inference services from July to November. The test reportedly caught at least three compromised hosting providers, resellers quietly logging prompts, and providers ignoring zero-data-retention flags. That flag is only a request, and the canaries showed who ignored it.

- **AI agents found a zero-day in 11 minutes** — [Source](<https://aiweekly.co/>) — Published 2026-11-21

Amazon says China-linked attackers used Cairn AI agents to find a zero-day in MONAI, an open-source framework for medical-imaging AI, in 11 minutes. That number is what should worry you. Most patch schedules assume attackers who need days or weeks to find the hole.

- **OpenAI says GPT-6.2 shuts down reasoning extraction** — [Source](<https://aiweekly.co/ai-news-today>) — Published 2026-11-21

Something for the defenders. OpenAI says new GPT-6.2 safeguards cut the success rate of reasoning-extraction attacks, which try to pry out a model's hidden reasoning, from 54% to 0.2%. That's OpenAI's own number, so outside red-teamers will want their turn.

- **Inkling 2 Mini, squeezed into 5.4 GB** — [Source](<https://github.com/ShinMegamiBoson/lookatmyquant>) — Published 2026-11-21

Thinking Machines opened the weights for Inkling 2 Mini, a 26.4B-parameter transformer that runs its middle layers three times in a loop. Within hours, the community had a quantized build (weights stored at lower precision) that reportedly fits in 5.4 GB of memory. Plenty of everyday machines can handle that.

- **Tencent opens up Hy4 flash weights** — [Source](<https://aitoolsrecap.com/Blog/ai-news-november-21-2026>) — Published 2026-11-21

Tencent released open weights for Hy4 flash: 224B parameters total, only 12B active at a time. That split usually means mixture-of-experts, where the model wakes up just a slice of itself for each step. Compute stays cheap, but your memory still has to hold all 224B.

- **AMD ships Nitro-P, XDNA driver hits 1.0** — [Source](<https://aiweekly.co/ai-news-today>) — Published 2026-11-21

Nitro-P is AMD's 679M-parameter text-to-image diffusion model. It works directly in pixels and skips the VAE, the compression step most image models rely on. Separately, AMD's XDNA driver for its NPUs, the dedicated AI chips in newer laptops, hit 1.0 for Linux 7.5. Anyone running it can exhale a little.

- **Genie 4 lands in the Gemini API** — [Source](<https://aitoolsrecap.com/Blog/ai-news-november-21-2026>) — Published 2026-11-21

Google's Genie 4 generates playable 720p worlds with audio, and builders can now call it straight from the Gemini API. That puts interactive, game-like scenes inside your own app. Expect your feed to fill up with gloriously weird prototypes by Monday.

- **Google traced GKE timeouts to a kernel bug** — [Source](<https://www.infoq.com/>) — Published 2026-11-21

Google's engineers chased API timeouts in GKE, their managed Kubernetes service, all the way down to a bug in the Linux kernel's TCP code. Every on-call engineer knows the prayer: please let it be DNS. This time the answer sat a few layers lower. Read it for the hunt.

- **OpenAI's Codex now tunes ChatGPT's performance** — [Source](<https://dailyaicaffeinenobs.substack.com/p/todays-real-ai-news-21-nov-26>) — Published 2026-11-21

OpenAI reportedly built a "Performance Factory," a set of Codex-powered workflows that find and fix performance problems in ChatGPT. So AI is tuning AI's own app. The part I'd pay to read is how they confirm each fix holds up under real production traffic.

- [An Agent's Identity Settles the Easiest of Three Questions](<https://current.tinyfish.ai/issue/55/foundations/article/65059/markdown>) — Published 2026-11-13

You ask an AI assistant to pay one invoice. A prompt appears, you tap approve, and it feels like you agreed to exactly that. But what did the tap actually authorize? A draft standard for governing AI agents offers an odd way into the question. It says it doesn't need to know whether the thing making the request is an agent. For a rule written about agents, that sounds like a gap. Follow it, though, and it leads back to your approve button.

- [The "I Agree" Button Assumed the Reader Was the One Clicking](<https://current.tinyfish.ai/issue/55/echoes/article/65167/markdown>) — Published 2026-11-13

In the 1990s, a court fight over a phone directory sold on CD-ROM helped set the rules for the "I agree" boxes people click online. Those rules decide when a click binds you to a seller's terms. Now AI shopping agents are doing the clicking for people. A shopper approves a printer, and the agent completes the purchase, but the seller's records show a printer and an arbitration clause. The shopper never saw that clause, yet under those rules they may still be bound by it.

- [What a 1984 Airline Screen Rule Would Ask of ChatGPT's Unpaid Answers](<https://current.tinyfish.ai/issue/55/market-pulse/article/65029/markdown>) — Published 2026-11-13

ChatGPT has started showing ads from businesses, and that brings back an argument the airline industry had in the 1980s. Back then, travel agents booked through reservation systems that airlines owned, and one of those systems found that 93 percent of agents' bookings came from the first screen of flights. Whatever appeared on that screen had a large advantage. This essay sets the two cases side by side and asks: if your business isn't among the advertisers, how would you know whether payment had shaped ChatGPT's ordinary, unpaid answers?

- [The Check-In Statement an Agent Can Carry but Not Make](<https://current.tinyfish.ai/issue/55/practitioners-corner/article/65149/markdown>) — Published 2026-11-13

United has moved to block AI agents from airline check-in, citing a box on the form that only a human can tick. Check-in has rested on the passenger's own word before. Until 2002, staff asked every passenger, "Has your bag been out of your control since you packed it?" Only the traveler could honestly answer. Once an agent is filling in the form for you, which answers can it give on your behalf, and which still have to come from you?

### [Market Pulse](<https://current.tinyfish.ai/issue/55/market-pulse/markdown>)

Reading the agent ecosystem through a practitioner's lens

- [What a 1984 Airline Screen Rule Would Ask of ChatGPT's Unpaid Answers](<https://current.tinyfish.ai/issue/55/market-pulse/article/65029/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-11-13 · Format: Market Moves

ChatGPT has started showing ads from businesses, and that brings back an argument the airline industry had in the 1980s. Back then, travel agents booked through reservation systems that airlines owned, and one of those systems found that 93 percent of agents' bookings came from the first screen of flights. Whatever appeared on that screen had a large advantage. This essay sets the two cases side by side and asks: if your business isn't among the advertisers, how would you know whether payment had shaped ChatGPT's ordinary, unpaid answers?

- **Developers pay model makers for every token processed** — [Source](<https://platform.claude.com/docs/en/build-with-claude/thinking-steering-and-cost>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) — Published 2026-11-13

Builders pay per million tokens, the word fragments a model reads and writes. Anthropic's new Fable and Mythos 5.2 cost **$8 in, $40 out**. Google's Gemini 4 Carbon costs **$2 and $10**, doubling after November 30. Anthropic bills a model's hidden reasoning as output.

- **Advertisers pay OpenAI per impression or valid click** — [Source](<https://help.openai.com/en/articles/20001651-setting-up-a-sponsored-business-agent-in-ads-manager>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) — Published 2026-11-13

Businesses can claim a ChatGPT profile and publish their own agent **free, for now**. OpenAI's paid product is ad delivery, billed by impression or valid click, with a Sponsored Business Agent as the ad's destination. Conversations with that agent and resulting sales carry no charge.

- **Enterprises buy IBM labor wrapped around prebuilt agents** — [Source](<https://newsroom.ibm.com/2026-11-10-ibm-and-microsoft-announce-a-new-ai-agent-platform-designed-for-industry-specific-transformation>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) — Published 2026-11-13

Imagine Enterprise pairs IBM Consulting's industry teams and **6,000-plus prebuilt agents and supporting assets** with Microsoft's cloud, data, and security. IBM can run it, build it with the client, or help clients manage it themselves. The launch discloses no rates, license fees, or revenue split.

- **Engineering teams pay LangChain per seat, then usage** — [Source](<https://www.langchain.com/pricing>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) — Published 2026-11-13

LangChain's open-source framework is free. The paid product is LangSmith, which monitors, tests, and deploys agents: **$39 per seat a month** on the Plus plan, including 10,000 traces (records of one full run each), then usage fees. The $3 billion valuation reflects investment, not sales.

- **Airbnb reports paying outside support contractors less** — [Source](<https://www.sec.gov/Archives/edgar/data/1559720/000155972026000032/abnb-20260930.htm>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) — Published 2026-11-13

Airbnb's quarterly filing links a **$26 million drop** in outside service costs, which include support contractors, to fewer contacts with human agents as AI took on more. About 40% of issues now resolve without a person. Total support spending still rose 8%.

- **SOPBench: Evaluating Language Agents at Following Standard Operating Procedures and Constraints** — [Source](<https://aclanthology.org/2026.emnlp-main.936.pdf>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

In seven simulated customer-service settings, agents were given tools, hints, the correct workflow order, and even the right verdict on whether each rule applied. The only call left was whether to act or refuse. Depending on setup, 12–38% of cases still went wrong.

- **The Hard Part Comes After Search: Benchmarking Web Agents on Synthesizing, Organizing, and Displaying Knowledge** — [Source](<https://aclanthology.org/2026.findings-emnlp.2127.pdf>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

KNOWS, a Findings of EMNLP 2026 benchmark, gave web agents 110 browser tasks that each end in a document, spreadsheet, or slide deck. The best system passed 64.7% of individual checks on average, yet fully finished only 2.7% of tasks.

- [Four Uber Rides Booked by AI Agents, and the Number Uber Left Out](<https://current.tinyfish.ai/issue/55/market-pulse/article/65024/markdown>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) · Published 2026-11-13

Uber says AI agents built by other companies booked [four rides](https://techcrunch.com/2026/11/04/almost-no-one-uses-ai-agents-to-book-their-uber-company-says/) during Halloween week. Uber handled roughly 320 million trips that same week. Plenty of people read that as proof nobody wants a bot hailing their car.

- **A CISO's caution on Microsoft's backdoor research** — [Source](<https://www.itpro.com/technology/artificial-intelligence/microsoft-says-small-tweaks-to-model-weights-could-turn-ai-into-sleeper-agents>) — Published 2026-11-13

Before agents scan models for hidden backdoors, Exabeam's CISO wants error rates and costs.

- **Menlo's caveat on its agent-usage numbers** — [Source](<https://menlovc.com/perspective/2026-the-state-of-enterprise-ai/>) — Published 2026-11-13

Agents' 80% share of tokens comes from OpenRouter, which Menlo says overrepresents startups and developers.

- **Users who'd rather plan their own fun** — [Source](<https://www.techradar.com/ai-platforms-assistants/uber-doordash-and-airbnb-say-ai-agents-arent-taking-off-as-users-rebel-against-being-hyper-efficient-at-fun>) — Published 2026-11-13

Some people like choosing dinner and trips themselves; others find assistants more work than apps.

- **Airbnb and DoorDash report thin agent use** — [Source](<https://www.techspot.com/news/114516-uber-doordash-airbnb-reveal-how-few-people-actually.html>) — Published 2026-11-13

DoorDash counts several hundred agent orders in total; Airbnb's Chesky says travelers like to browse.

- [When AI Models Get Cheaper, Who Keeps the Savings?](<https://current.tinyfish.ai/issue/55/market-pulse/article/61577/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-11-05 · Format: Market Moves

When a customer pays SAP a fixed price for each task SAP's AI handles, the model doing the work can change while the price stays put. SAP is making that swap, moving work onto cheaper models and keeping the difference. In a written answer to Reuters, it said so plainly: "The goal is healthy margins as scale grows, not an absolute reduction in AI expenditure."

- [Who Pays While the Agent Waits](<https://current.tinyfish.ai/issue/55/market-pulse/article/57577/markdown>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) · Published 2026-10-30 · Format: Market Moves

Google's new service for hosting AI agents starts from a fact anyone who has watched one run already knew: most of the time, the agent is waiting, for the model to answer or for a person to reply. So Google suspends idle agents and promises you won't pay for idle machines. Fair enough, but the wait still gets paid for. It moves into approval queues nobody times, and then onto the plan, which comes back exactly as it left, even if the stock it counted on has since sold.

- [Where Inertia Goes When Agents Do the Checking](<https://current.tinyfish.ai/issue/55/market-pulse/article/52865/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · 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.

- [The Affected Non-Participant](<https://current.tinyfish.ai/issue/55/market-pulse/article/45883/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · 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.

### [Foundations](<https://current.tinyfish.ai/issue/55/foundations/markdown>)

Conceptual clarity earned from building at scale

- [An Agent's Identity Settles the Easiest of Three Questions](<https://current.tinyfish.ai/issue/55/foundations/article/65059/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-11-13 · Format: Core Concepts

You ask an AI assistant to pay one invoice. A prompt appears, you tap approve, and it feels like you agreed to exactly that. But what did the tap actually authorize? A draft standard for governing AI agents offers an odd way into the question. It says it doesn't need to know whether the thing making the request is an agent. For a rule written about agents, that sounds like a gap. Follow it, though, and it leads back to your approve button.

- [A Timed-Out Payment Is Unknown, Not Failed](<https://current.tinyfish.ai/issue/55/foundations/article/65079/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-11-13 · Format: Failure Walkthrough

An AI agent asks Stripe to charge a customer's card, and Stripe does it: the card is charged, and Stripe sends back the ID of the new payment. But the connection drops before that reply reaches the agent. The agent waits, raises an error, and its code files the order as declined. The customer has paid, yet the agent's records say the charge never happened. What follows is a walkthrough of how that happens and why your own records end up disagreeing with Stripe's.

- **Mapping prompt injections hidden in web pages** — [Source](<https://scnps.co/papers/ccs2026_prompt_injection.pdf>) — Published 2026-11-13

The crawl misses logged-in pages and agent-only content, yet found 15.3K injections in 1.2B URLs.

- **A debugging tool that never timed developers** — [Source](<https://aclanthology.org/2026.emnlp-demos.52.pdf>) — Published 2026-11-13

The authors never measured developer time; the tool recovered 13 of 73 benchmark failures.

- **AI judges scored by a flawed answer key** — [Source](<https://aclanthology.org/2026.emnlp-main.659.pdf>) — Published 2026-11-13

Its answer key marks down valid extra tool-call arguments, and one annotator checked each record.

- **Retrieval benchmark too small for close calls** — [Source](<https://aclanthology.org/2026.aacl-long.245.pdf>) — Published 2026-11-13

With about 60 questions per category, the authors warn, small gaps between models stay unsettled.

- **Routing tasks when an agent abstains** — [Source](<https://aclanthology.org/2026.realm-1.57.pdf>) — Published 2026-11-13

Evidence comes from one e-commerce assistant, where task success rose from 71.5% to 84%.

- **Surveying deployed agents, with a sampling caveat** — [Source](<https://proceedings.mlr.press/v306/pan26a.html>) — Published 2026-11-13

The authors flag sampling bias in their 86 deployed systems, where reliability topped reported problems.

- **Agent says done, customer got nothing** — [Source](<https://github.com/open-telemetry/semantic-conventions-genai/blob/main/docs/gen-ai/mcp.md>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

Start at the destination system and work backward to the agent's summary. Compare what the agent claimed, what the tool returned, and what the target system recorded. A tool call over MCP, a common agent-to-tool protocol, can finish cleanly and still carry `isError: true`. If a write may have landed before a timeout, see this issue's payment walkthrough.

- **Same input, different result on every rerun** — [Source](<https://aclanthology.org/2025.eval4nlp-1.12.pdf>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

The user's message is a small slice of what the model actually received. Diff the instructions, model settings, retrieved documents, tool definitions, and memory between runs. One study saw five hosted models vary across ten runs at supposedly deterministic settings. Another saw rankings move when only the date in the system prompt changed.

- **The answer's in our docs. Still wrong.** — [Source](<https://papers.neurips.cc/paper_files/paper/2024/file/27245589131d17368cccdfa990cbf16e-Paper-Datasets_and_Benchmarks_Track.pdf>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

Split it into two questions. Did the fact you needed reach the model's context at all? If it did, did the answer actually use it? RAGChecker, an evaluation framework for retrieval systems, scores those separately. It also found that pulling in more context can just hand the model more noise to trip over.

- **A one-time exception became a permanent rule** — [Source](<https://aclanthology.org/2026.findings-emnlp.571.pdf>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

Follow the memory from start to finish: the original conversation, what got saved, what type it was filed as, and where it came back. The MemGuard paper documents passing events hardening into durable "facts," and anecdotes overriding explicit rules, when stored facts, events, and rules get mixed together.

- **It sees the button but can't click** — [Source](<https://www.w3.org/TR/2026/WD-webdriver2-20261109/>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

Save the exact browser error, the frame, the element reference, and a screenshot. The WebDriver automation standard separates three failures: the element went stale after the page changed, something else is covering it, or it can't be interacted with where it sits. A delivered click also doesn't prove it hit the right button.

- **Works for the developer, fails for customers** — [Source](<https://www.rfc-editor.org/rfc/rfc6750.html#section-3.1>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

Check which credential each run actually used, since a developer's own token may carry permissions a customer's doesn't. OAuth separates an invalid token (expired, revoked, malformed: HTTP 401) from insufficient scope (HTTP 403). Under MCP, missing scope can hide a tool entirely, so the agent never even attempts the call.

- **It keeps doing the same thing, forever** — [Source](<https://proceedings.nips.cc/paper_files/paper/2025/file/b1041e52d3be19f0a9bc491657488e4a-Paper-Datasets_and_Benchmarks_Track.pdf>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

Line up each action next to what the agent observed, then check what state each step could actually see. The MAST study of multi-agent failures separates repeated steps, lost history, missed stopping points, and information one agent never passed on. In one trace, an unshared username format caused endless failed logins.

- **It stops halfway through, with no crash** — [Source](<https://github.com/open-telemetry/semantic-conventions-genai/blob/main/docs/gen-ai/client-inference.md>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

Grab the model provider's finish reason (`stop`, `length`, `error`), the configured output limit, and the partial response. Text cut off at `length` can read a lot like a finished answer. Then check the agent runtime's own record of why it decided the task was over.

- **The query runs. The answer is wrong.** — [Source](<https://aclanthology.org/2026.emnlp-industry.42.pdf>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

Running is a low bar. On FLINT's production financial datasets, seven text-to-SQL methods wrote queries that executed over 99% of the time, and none topped 50% accuracy. Check which business definition the agent pulled, which reference values it resolved, and whether it applied them correctly.

- **Revision 2026-07-28 is published and current** — [Source](<https://modelcontextprotocol.io/specification/2026-07-28/server/tools>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

MCP (Model Context Protocol) is how many agents call external tools. Its July 28 release is the current revision, and tools are a section of it. If a tool declares an output schema, servers **must** return conforming results and clients **should** validate them. *Current* still allows compatible changes.

- **Watch: which revision your client actually negotiates** — [Source](<https://github.com/modelcontextprotocol/modelcontextprotocol/blob/main/docs/docs/2026-07-28/learn/versioning.mdx>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

Publication doesn't tell you what your client speaks. Requests declare a protocol version, servers report the versions they support, and both sides can run several at once. Check what you're actually negotiating. Tasks also moved out of core into an extension with its own status.

- **Community draft, with browser origin trials running** — [Source](<https://webmachinelearning.github.io/webmcp/>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

WebMCP lets a web page expose JavaScript functions as tools an agent can call. It's a Draft Community Group Report, not a W3C Standard. Chrome 149 and Edge 150 list origin trials (time-limited opt-in tests for sites), plus a local-dev flag. Experiment, and expect the spec to move.

- **Watch: who validates output schemas, and how strictly** — [Source](<https://github.com/webmachinelearning/webmcp/issues/92>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

Output schemas are optional today, and agents may reject results that don't match. Issue #92 records a March resolution to put validation in the browser, but the supported JSON Schema subset is undefined and contributors disagree on making schemas mandatory. Watch for actual spec text.

- **Revision 02: an individual draft without IETF standing** — [Source](<https://datatracker.ietf.org/doc/draft-rosenberg-oauth-aauth/>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

AAuth is a proposed protocol for authorizing AI agents. Revision 02, by Jonathan Rosenberg and Pat White, was posted November 1, 2026. Datatracker lists it as an active individual Internet-Draft, which the IETF does not endorse. The Standards Track label in its header is a goal, not a status.

- **Watch: renewal before the May 5, 2027 expiry** — [Source](<https://datatracker.ietf.org/doc/draft-rosenberg-oauth-aauth/history/>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

Internet-Drafts expire after six months unless replaced. Revision 01 expired April 22, and 02 didn't appear until November, so gaps happen. Revision 02 expires May 5, 2027. Watch the history for renewal or lapse, and the status record for any formal change, such as working-group adoption.

- **allowed-tools is optional, experimental, and unevenly supported** — [Source](<https://agentskills.io/specification>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

A skill is a folder of instructions and metadata, sometimes with scripts, that an agent loads when relevant. The optional `allowed-tools` field lists tools pre-approved to run, as a space-separated string. The spec marks it experimental and says support varies by agent. Confirm your host honors it.

- **Watch: string or list, or outright replacement** — [Source](<https://github.com/agentskills/agentskills/issues/144>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

Issue #144 debates whether the field should stay a string or become a YAML list. One contributor quotes project discussion suggesting the field, inherited from the pre-standard format, could give way to a vendor-neutral mechanism. Nothing is adopted. For now, your host's own docs decide what actually happens.

- [Test a Page's Agent Tools the Way You Test Its Buttons](<https://current.tinyfish.ai/issue/55/foundations/article/61611/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-11-05 · Format: Practical Walkthrough

Web pages are starting to hand AI agents tools instead of buttons. Rather than clicking, the agent calls a tool, sends it some inputs, and gets something back. That raises a question clicking never did: who checks what goes in and what comes out?

- [Revenue Is Number 28, and Someone Has to Tell the SQL Agent](<https://current.tinyfish.ai/issue/55/foundations/article/61610/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-11-05 · Format: Core Concepts

Ask an AI agent a question about your company's data, and it writes a database query, runs it, and returns a number. New research tested how often that number is right. Running was rarely the problem: every system tested wrote queries that ran at least 99% of the time. Yet apart from the researchers' own system, none answered even half the questions correctly. The wrong answers look as exact as the right ones, and nothing warns you which is which. This close read asks what a query that runs actually tells you.

- [Three Ways an Agent Can See a Page, and What Each One Drops](<https://current.tinyfish.ai/issue/55/foundations/article/57538/markdown>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) · Published 2026-10-30 · Format: Hands-On Walkthrough

Before a web agent clicks anything, somebody has already decided which version of the page it gets. The usual choices are a screenshot, the accessibility tree browsers build for screen readers, or the DOM, the browser's full model of the page. Each one makes some questions easy and others impossible.

- [Your Agent's Questions Need a Recipient and a Timer](<https://current.tinyfish.ai/issue/55/foundations/article/57518/markdown>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) · Published 2026-10-30 · Format: Core Concepts

When an AI agent stops mid-task to ask a human something, it often reads as a failure of autonomy. In one study, letting coding agents ask about bug reports stripped of key details won back most of the lost ground, lifting solve rates from about 55% to 69.4%. But the "human" answering was another model, holding the full report and standing by. Real people are in meetings, or asleep. The design work is in the ask: who gets it, what it says, and what happens while it waits.

### [Practitioner's Corner](<https://current.tinyfish.ai/issue/55/practitioners-corner/markdown>)

Lessons from the field—what we see building at scale

- [Melissa Pan and the Logs the Field Can't See](<https://current.tinyfish.ai/issue/55/practitioners-corner/article/65150/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-11-13 · Format: In Practice

Pan began as an engineer recovering databases after they failed. Now she studies why AI agents fail, but the companies running them keep their records private, so the field learns from whatever is public. Pan once learned that way too: as a grad student, she has written, she "had the view that most agents were coding agents," the kind most visible on public leaderboards, until production data corrected her. If the public record misled her about what most agents do, what might it miss about why they fail?

- [The Check-In Statement an Agent Can Carry but Not Make](<https://current.tinyfish.ai/issue/55/practitioners-corner/article/65149/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-11-13 · Format: Limits of Delegation

United has moved to block AI agents from airline check-in, citing a box on the form that only a human can tick. Check-in has rested on the passenger's own word before. Until 2002, staff asked every passenger, "Has your bag been out of your control since you packed it?" Only the traveler could honestly answer. Once an agent is filling in the form for you, which answers can it give on your behalf, and which still have to come from you?

- [Walking the Board at 3 A.M., an Interview With the Night Auditor of a Browser Fleet](<https://current.tinyfish.ai/issue/55/practitioners-corner/article/65135/markdown>) — Published 2026-11-13

Every team that runs a browser fleet eventually says "we keep the sessions alive." They usually say it to a manager, and they usually say it with confidence. It sounds like a single promise. Read the specifications underneath and it comes apart into at least four: the automation's connection to the browser, the pages that are open, the data the browser has squirreled away, and the website's private opinion about whether you're still allowed in. Each has a different owner and runs on a different clock. The W3C's WebDriver draft, revised again this Monday, has a careful definition of "session." It covers the first one.

- [Railway Rotated a Key and Its Backups Quietly Stopped](<https://current.tinyfish.ai/issue/55/practitioners-corner/article/65136/markdown>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) · Published 2026-11-13

Railway's [write-up of its October security incident](https://blog.railway.com/p/incident-report-october-security-incident) has one line I can't stop chewing on. During cleanup, the team rotated the credentials its services use to write to storage. That was the correct move. Then database backups quietly stopped landing in storage. They didn't resume until the affected services were redeployed.

- **Bridge Law 16: Information You Can't Unsee** — [Source](<https://www.worldbridge.org/wp-content/uploads/2019/01/2017LawsCommentary.pdf>) — Published 2026-11-13

A withdrawn bid or card leaves information that only the opponents may use.

- **Pacioli's Renaissance Method for Fixing Entries** — [Source](<https://gwern.net/doc/bitcoin/1963-brown-pacioloonaccounting.pdf>) — Published 2026-11-13

Post an offsetting entry, explain the mistake, then record it correctly, leaving everything visible.

- **Reversing an Accrual, Using Copy Paper** — [Source](<https://www.cu.edu/controller/fye-accrual-example>) — Published 2026-11-13

A university controller's example shows how a reversal stops one expense being counted twice.

- **Livent: Fraud by Erasing the Original Entry** — [Source](<https://www.sec.gov/enforcement-litigation/administrative-proceedings/34-40958>) — Published 2026-11-13

SEC findings show how deleting original entries left auditors no trail of the adjustments.

- **Airfare Refunds: Who Owes What, When** — [Source](<https://www.transportation.gov/individuals/aviation-consumer-protection/refunds>) — Published 2026-11-13

Federal rules assign refund duty based on timing, cause, and who actually sold the ticket.

- **What Reversibility Means in Art Conservation** — [Source](<https://cool.culturalheritage.org/jaic/articles/jaic26-02-001.html>) — Published 2026-11-13

A conservator argues a reversible treatment is one that keeps later treatment options open.

- [Daniel Stenberg Named the Cost Before Asking for Help](<https://current.tinyfish.ai/issue/55/practitioners-corner/article/61484/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-11-05 · Format: Builder Profile

The code in curl runs in billions of devices, so when someone reports a security flaw in it, its maintainer has to find out whether the flaw is real. AI tools have made such reports cheap to send, but each one is still slow to check. By May, more than one report a day was arriving, nearly all of them AI-assisted, about twice the 2025 rate. This profile follows a maintainer under that pile who is betting that strangers will step in to help.

- [Phantom Stock and the Checks That Never Fire](<https://current.tinyfish.ai/issue/55/practitioners-corner/article/61483/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-11-05 · Format: In Practice

A grocery store's computer says the shelf holds six cans of tomatoes. The shelf is empty. Because the computer believes the cans are there, it never orders more, and the shelf stays empty while the records say all is well.

- [PHP's Memory Bug Had Fixes Long Before It Had a Decision](<https://current.tinyfish.ai/issue/55/practitioners-corner/article/57552/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-10-30 · Format: Feature Essay

On October 21, PHP, the language running roughly seven in ten websites whose server language can be identified, shipped emergency fixes for every version it maintains. One flaw let an anonymous outsider run code on some servers by making PHP run out of memory at the wrong moment.

- [Faria Huq and the Reason Behind a Human Step](<https://current.tinyfish.ai/issue/55/practitioners-corner/article/57548/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-10-30 · Format: In Practice

When a person takes control from an AI agent driving their web browser, the system records a human step, but not why. Faria Huq and her Carnegie Mellon collaborators have built tools around these moments: an agent that can't find the right forum, a shopper who realized the brand mattered only after seeing the agent's pick. Their system learns to predict when someone will step in. The reason it sets aside would tell a team whether to fix the agent, the website, or its understanding of the person.

### [Echoes](<https://current.tinyfish.ai/issue/55/echoes/markdown>)

Past infrastructure decisions echoing in today's production realities

- [The "I Agree" Button Assumed the Reader Was the One Clicking](<https://current.tinyfish.ai/issue/55/echoes/article/65167/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-11-13 · Format: Historical Parallels

In the 1990s, a court fight over a phone directory sold on CD-ROM helped set the rules for the "I agree" boxes people click online. Those rules decide when a click binds you to a seller's terms. Now AI shopping agents are doing the clicking for people. A shopper approves a printer, and the agent completes the purchase, but the seller's records show a printer and an arbitration clause. The shopper never saw that clause, yet under those rules they may still be bound by it.

- [The Bracket Unicode Can Never Spell Right](<https://current.tinyfish.ai/issue/55/echoes/article/65160/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-11-13

In 2005, Unicode, the standard that gives every character a number and a permanent name, added sideways brackets for Chinese written top to bottom. One shipped as PRESENTATION FORM FOR VERTICAL RIGHT WHITE LENTICULAR **BRAKCET**.

- [In Conversation With the Courier Who Held Everything and Owned Nothing](<https://current.tinyfish.ai/issue/55/echoes/article/65154/markdown>) — Published 2026-11-13

In 1990, a roomful of maritime lawyers in Paris wrote down something every cargo office already knew. Ships were reaching port before the paperwork that entitled anyone to unload them. Cargo was being handed over against letters of guarantee while the bill of lading, the document that says who gets the goods, was still in a courier's bag somewhere between banks. Their fix was an electronic bill of lading, with the carrier keeping score of who held it.

- [Java's Ten-Year Search for the Last Applets](<https://current.tinyfish.ai/issue/55/echoes/article/65174/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-11-13 · Format: Retiring Two Platforms

By 2017, browsers had stopped running applets, the small Java programs that once ran inside web pages. Removing applets from Java itself still took a decade, and much of that time went into finding everything else that had quietly been built on them. Some developers had borrowed the applet sound player just to make ordinary desktop programs beep, so Java had to ship a replacement before it could delete the original.

- [Flash Shipped With Its Own Shutdown Date](<https://current.tinyfish.ai/issue/55/echoes/article/65175/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-11-13 · Format: Retiring Two Platforms

Flash, the browser add-on behind a generation of web games, got a firmer ending. Adobe gave years of notice, then wrote a fixed quitting day into the player itself, so Flash would stop on schedule whether anyone was ready or not. That left everyone else to discover what still depended on it. In January 2021, South Africa's tax authority publicly admitted: "We erred in our interpretation that the functionality would continue beyond the support date."

- **Shannon's 1948 paper sets meaning aside** — [Source](<https://academicweb.nd.edu/~powers/ame.20231/shannon1948a.pdf>) — Published 2026-11-13

Messages have meaning, he grants, then: “These semantic aspects … are irrelevant to the engineering problem.”

- **Where IBM's System/360 drew the compatibility line** — [Source](<https://people.cs.umass.edu/~emery/classes/cmpsci691st/readings/Arch/Architecture-of-the-IBM-System-360.pdf>) — Published 2026-11-13

Its architecture covered only “the attributes of a system as seen by the programmer.”

- **The 1988 RAID paper starts with bad news** — [Source](<https://www.cs.cmu.edu/~15712/papers/patterson88.pdf>) — Published 2026-11-13

“Without fault tolerance, large arrays of inexpensive disks are too unreliable to be useful.”

- **MIL-STD-105E, a retired military sampling standard** — [Source](<https://www.thomasmechanicaldesign.com/uploads/5/5/8/1/55818795/mil-std-105e.pdf>) — Published 2026-11-13

An allowed defect rate, it warns, grants no right “to knowingly supply any defective unit.”

- **RFC 2119, the memo that defined MUST** — [Source](<https://www.rfc-editor.org/rfc/rfc2119.html#section-6>) — Published 2026-11-13

Its own advice on those capitalized words: they “must be used with care and sparingly.”

- **How internet standards define rough consensus** — [Source](<https://www.rfc-editor.org/rfc/rfc7282.html#section-3>) — Published 2026-11-13

RFC 7282: “Rough consensus is achieved when all issues are addressed, but not necessarily accommodated.”

- [Screen Reader Users Never Took the Page at Its Word](<https://current.tinyfish.ai/issue/55/echoes/article/61529/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-11-05 · Format: Paired Stories

On the web, code written for blind users is picking up visitors it was never built for. Hidden labels in a page tell blind users' software "this is a button," and those users learned the labels don't always deliver. The web standards body's guide calls such a label a role and warns developers, "A role is a promise," because tagging a plain box as a button doesn't make it work like one. Now AI agents that use websites on people's behalf are reading those same labels.

- [Mobile Money Still Runs Inside the Operator's Session](<https://current.tinyfish.ai/issue/55/echoes/article/61528/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-11-05 · Format: Paired Stories

In Nigeria and Kenya, money moves through text menus on ordinary phones, using a 1990s design that was never meant to carry it. That design assumed the mobile network writing a menu also owned the call the menu ran over. Then banks started writing the menus, while the networks kept the calls. In October 2019, Nigeria's largest network announced a charge of ₦4 for every 20 seconds a customer spent in a bank's menu, metering money transfers the way it meters phone calls.

- [Why QUIC Shipped a Version 2 With No New Features](<https://current.tinyfish.ai/issue/55/echoes/article/61520/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-11-05 · Format: Historical Parallels

Internet protocols were built to be upgradable. In 2017, a Chrome update tested that in Montgomery County, Maryland, where upwards of 30 percent of the district's roughly 50,000 school Chromebooks got stuck looping at the login screen, locking students out.

- [Western Union Sold Accuracy as an Upgrade](<https://current.tinyfish.ai/issue/55/echoes/article/57662/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-10-30 · Format: Paired Features

In 1887 a Philadelphia wool dealer wired his Kansas buyer "bay," code for "I have bought." It arrived as "buy." The buyer bought, the loss ran past $20,000, and the Supreme Court held that Western Union owed $1.15, the price of the telegram. Accuracy had been an upgrade: having the message read back cost under sixty cents more. Software placing orders on other people's websites still faces the question of who pays to confirm a message arrived as sent.

### [Vision](<https://current.tinyfish.ai/issue/55/vision/markdown>)

Where human-AI collaboration is heading

- [Speaking to the Reader Who Isn't on the Call](<https://current.tinyfish.ai/issue/55/vision/article/65103/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-11-13 · Format: Possible Futures

A director's colleague couldn't join client calls, so he spent five minutes listing ten concerns, and the director carried that list in and said things aloud on purpose so the colleague could question the transcript afterward. It's a sensible workaround. It also means the calls had a listener the client may not have known about. AI now lets colleagues who weren't in a meeting search what was said later. What happens to candid conversation when only one side may know who's listening?

- [The Human Backup Needs a Drill](<https://current.tinyfish.ai/issue/55/vision/article/65105/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-11-13 · Format: Expertise and Roles

AI now does much of the everyday work that keeps experts sharp. Companies count on whoever signs off on that work to catch the AI's mistakes, yet rarely test whether they still can, especially when a wrong answer looks right. Better output doesn't prove the skill is there. In a study of 487 management consultants trying unfamiliar data tasks, those using AI beat colleagues without it by 49 percentage points on one task. On a short quiz afterward without the AI, both groups scored about the same.

- [The Apprenticeship Clients Used to Fund](<https://current.tinyfish.ai/issue/55/vision/article/65106/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-11-13 · Format: Expertise and Roles

That same everyday work is also how people become experts. Young accountants and associates learned their trade on routine tasks that clients paid for, and AI now does those tasks. A Stanford study finds that employment of 22-to-25-year-olds in the jobs most exposed to AI has fallen 19 percent behind their peers in other jobs. The main reason is that firms are hiring fewer of them, though the authors don't blame AI alone. The missing training may not show until today's senior people retire.

- [An Interview With Prudence Haircut, the Bank Cost Manager Who Has to Price Checking the Machine](<https://current.tinyfish.ai/issue/55/vision/article/65094/markdown>) — Published 2026-11-13

On Friday, November 13, three days after Reuters reported that banks are handing their AI bills to new teams some call **AI FinOps**, we sat down with someone whose job barely has a description yet. (FinOps is the discipline that grew up around taming cloud bills. Its newest assignment is the meter running on AI agents.)

- [Do It, Show Me, or Help Me Decide](<https://current.tinyfish.ai/issue/55/vision/article/65095/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-11-13

At EMNLP, a language-technology conference, researchers last month described [eight blind people](https://aclanthology.org/2026.emnlp-main.1763.pdf) who spent three weeks working with an AI agent that could operate twelve desktop apps for them. Sometimes they asked it to finish a job. They also asked where a menu was and what an unfamiliar screen meant, so they could manage alone next time.

- **Web Standards Body Drafts Agent Explainers** — [Source](<https://www.w3.org/2026/11/12-webmachinelearning-minutes.html>) — Published 2026-11-13

Due by month-end, they show which agent identity and delegation problems the W3C might tackle.

- **Tracking a Proposed W3C Agent Standards Group** — [Source](<https://github.com/w3c/strategy/issues/576>) — Published 2026-11-13

Its charter cleared security review; this tracker will flag when members are asked to weigh in.

- **United Restricts Third-Party AI Booking Agents** — [Source](<https://united.mediaroom.com/news-releases?item=125496>) — Published 2026-11-13

Watch the promised policy updates and help pages to see how far the restrictions reach.

- **Why Delta and American Aren't Following United** — [Source](<https://www.fastcompany.com/91640548/united-airlines-just-made-a-big-change-to-flight-bookings-heres-why-delta-and-american-arent-following>) — Published 2026-11-13

Both accept some authenticated agents but leave passengers responsible for what those agents do.

- **UK Finance Regulator Weighs Agent Accountability** — [Source](<https://www.fca.org.uk/publication/corporate/apm-2026-responses-unanswered-questions.pdf>) — Published 2026-11-13

The FCA is considering clearer rules on complaints and redress when consumers use AI agents.

- **Four UK Regulators Set 2027/28 Priorities** — [Source](<https://www.drcf.org.uk/siteassets/drcf/home/newsevents/drcf-2027-28-workplan-cfi.pdf?v=426511>) — Published 2026-11-13

Submissions close November 20 and could put agent redress on their shared agenda.

- [Selling to a Shopper That Reasons](<https://current.tinyfish.ai/issue/55/vision/article/61663/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-11-05 · Format: Paired Features

Shopping has changed who does the choosing before. In 1916, Piggly Wiggly in Memphis let shoppers take goods off the shelf themselves instead of handing a list to a clerk, so the box had to do the clerk's selling. Now shoppers are handing the list back, this time to an AI agent that chooses for them, and sellers are already rewriting their product information for it. The open question is whether a smarter agent resists that pitch.

- [The Agent Your Agent Hired](<https://current.tinyfish.ai/issue/55/vision/article/61662/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-11-05 · Format: Paired Features

Letting an AI agent act for you repeats an old arrangement: someone doing a job in your name. Now agents are starting to pass your tasks to other agents you never picked, and centuries of law on acting for someone else ask who is really deciding. Even taking your permission back may not settle it. Starting November 24, Wells Fargo will ask customers who cancel an agent's access to their accounts to take steps to stop it, and to keep proof they did.

- [When Agents Use the Doorway Built for Blind Users](<https://current.tinyfish.ai/issue/55/vision/article/61648/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-11-05 · Format: Possible Futures

When a developer forgets to give a button a name, a blind person using a screen reader hears only one word, "button," and nothing about what it does. Blind users have put up with this for years. Now AI agents reading the same page are left guessing too. That gives the companies paying for those agents a reason to finally fix the missing labels, helping blind users along the way. Or those companies could build their agents a way around the problem and leave the buttons unnamed.

- [The Interviewer Worked for the Firm](<https://current.tinyfish.ai/issue/55/vision/article/57489/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-10-30 · Format: Paired Features

A job-interview bot owes the applicant none of the loyalty the W3C asks of user agents. The employer buys it and writes its questions. Yet in a new, not-yet-reviewed experiment covering about 67,000 call-center applications in the Philippines, applicants interviewed by the AI got more offers, started more often and stayed longer. The reason lies in what recruiters got to know, and it suggests a test for any go-between a firm deploys.

## More

- [Publication guide](<https://current.tinyfish.ai/llms.txt>)
- [Issue archive](<https://current.tinyfish.ai/issues/markdown>)
