mAIn Street #337 | Full Edition: Claude Duet, Walmart shopping agents, small-business AI and Barclays jobs
|
| FRIDAY, OCTOBER 23, 2026 · ISSUE #337 · FULL EDITION |
|
START HERE
What are we ready to hand over?
Claude is offering to keep track of work between prompts. Walmart is preparing to let agents shop under spending limits. Goldman Sachs is putting resources into small-business AI training, while Barclays is cutting jobs as part of a technology-driven cost effort.
Those are the four leads today. There is plenty more below, from a browser shutdown deadline and translated meetings to a hospital that found an AI pilot created too much checking. We finish with a practical Friday exercise for learning one small skill well enough to use it without the chat open.
|
|
|
01 · THE LEAD STORIES
Claude Duet will keep working after you close the chat
An assistant that notices an overdue invoice before you remember to ask could be useful to a busy owner. Anthropic’s new Claude Duet is designed to monitor connected apps, prepare work and flag matters needing attention. It can continue on its own cloud computer while your laptop is off.
The research preview starts with a waitlist for U.S. Pro and Max subscribers. Anthropic says Duet reads and drafts on its own, but asks before sending messages, making payments, signing documents or deleting things. Those are the company’s stated safeguards, not a reason to skip reviewing a proposed action.
There is a cost to that convenience even without a separate subscription charge: background checks and tasks use the same limits as the rest of your Claude plan. Anthropic also warns that the early version can misread a conversation or get urgency wrong.
If you get access, start with one low-risk job and a clear boundary: watch for unpaid invoices, prepare a list and do not contact customers. Check the usefulness of the list and the usage it consumes before handing over more work.
|
|
|
02 · THE LEAD STORIES
Walmart is putting spending controls around AI shopping
Letting an assistant buy the usual groceries sounds convenient until it buys the wrong size, brand or quantity. Walmart’s announced Agent Profiles will let shoppers separate agent purchases from their own and choose whether every purchase needs approval or falls under a spending limit.
The profiles are due within a month of the October 21 announcement. A separate connection for agent companies and developers starts in beta. Walmart says agents will see the same product prices as human shoppers, with applicable offers and fees disclosed, and its returns process will still apply.
The controls matter more than how smoothly the shopping conversation goes. A household or small office could use an agent for repeat supplies, but should still specify acceptable substitutions and review the total. A spending cap prevents some expensive mistakes; it does not tell an assistant which replacement you actually want.
For now, this is a rollout to watch rather than a feature every customer can switch on today. When it reaches your account, begin with approval required for each order.
|
|
|
03 · THE LEAD STORIES
Goldman Sachs puts small-business AI training on the agenda
Goldman Sachs announced a Small Business AI Institute as part of a broader $500 million commitment. The practical issue for owners is familiar: choosing an appropriate tool and making it useful can take more work than buying a subscription.
The accompanying survey found that 73% of AI-using alumni spend less than $100 a month, including free tools. The sample comprised 2,473 program alumni and may not represent all alumni, much less every U.S. small business.
The takeaway for a small operation is to price the whole experiment. Include the subscription, setup time and checking someone must do before an answer reaches a customer. Pick one recurring task, record how long it currently takes, then compare the finished work after a trial.
The institute’s announced plans include learning and workshops with Anthropic, Google, Meta and OpenAI. The broader funding also supports education, capital and other help for another 10,000 businesses. Check enrollment details: the announcement does not give every U.S. owner immediate free access.
|
|
|
04 · THE LEAD STORIES
Barclays’ job cuts deserve a closer look than “AI replaced everyone”
Barclays plans to cut roughly 5,000 jobs during 2026, about 5% of its workforce, as it pursues efficiencies involving AI and other technology. AFP reports that support functions are a major part of the reductions.
About 1,000 cuts had already happened in the first half of the year, according to the bank’s reported explanation. The total belongs to a longer modernization effort. It should not be presented as 5,000 new layoffs caused solely by a chatbot.
That distinction offers little comfort to someone whose position is affected. It does matter when deciding what the announcement says about everyone else’s work. A company’s cost-cutting plan is evidence about that company, not a reliable forecast for an entire occupation.
For workers, a useful conversation with a manager is specific: which tasks are changing, what training is available, and how will good work be judged? Keep examples of the decisions, checks and customer outcomes you contribute. AI proficiency may help with changing work, but no course or prompt can guarantee job security.
|
|
|
THE ROUNDUP
Using AI at work
Uber expands ChatGPT Work
Uber is putting 30,000 employees on ChatGPT Work after a 1,000-person pilot, mostly involving nonengineers.
Google Meet expands speech translation to 63 languages
Google’s rollout lets participants choose their listening language while Meet detects changes in what speakers use. It began October 22 for Rapid Release; Scheduled Release starts November 2. An eligible paid host is required, and meeting-duration and regional limits apply. A customer call could benefit, but check account access before promising translated meetings.
Arc users have a December deadline to move
The Browser Company says Arc’s desktop browser will stop working December 22. Mobile AI features switch off November 23. Users can export their data to another browser rather than move to the company’s AI-focused Dia. If Arc is part of your workday, schedule the move while you can still check that important tabs and information transferred.
1Password introduces a login layer for personal AI agents
1Password for Agents is designed to let users approve an assistant’s login without exposing the password to the model. Integrations are at different stages; Claude Duet support is described as coming soon. Keeping a password away from the model is useful, but users still need boundaries on what the agent may do after signing in.
AI bills are prompting employers to ration access
Korn Ferry reports that some companies are extending usage caps beyond developers to other workers. Its account quotes its own experts and unnamed employers, rather than measuring how common caps are. For a small team, review spending by task and useful outcome before imposing a blanket limit that interrupts valuable work.
CoCounsel puts calculations and workpapers beside tax answers
Thomson Reuters’ update emphasizes formulas, assumptions and supporting workpapers in professional tax and accounting workflows. The U.S. offering is available, but the announcement supplies no public price. Showing work can help a professional review it; the company’s accuracy language is not a guarantee that every answer is correct.
Haystack expands writing through AI assistants
Haystack says employees will be able to create intranet pages and posts through Claude, ChatGPT, Copilot Studio and Slack starting October 26. The cross-platform expansion requires a separate add-on. Writing through Claude existed previously, so the new development is the broader set of connections.
|
|
|
THE ROUNDUP
Jobs, training and small-business operations
Pew finds unequal job growth in AI-exposed occupations
Pew’s analysis finds stronger employment growth for men than women in occupations with greater potential AI exposure, with a wider gap among people without bachelor’s degrees. Exposure measures tasks AI might assist with or perform, not actual adoption. The association does not prove AI caused the gap.
Libraries get support for practical AI training
The Public Library Association and Microsoft announced free training and resources for library staff and patrons. The first webinar is December 3, followed by workshops into spring 2027. Local participation will vary, so this is a program to ask about rather than an assumption that every library already has classes.
CompTIA plans a certification for running AI agents
The planned AgentOps+ credential covers deployment, permissions, monitoring and human oversight for existing IT workers. Certification and learning products are expected in the second quarter of 2027. It is an announced training path, with no price or exact exam date yet, rather than a credential readers can earn now.
Restaurants add AI support for delivery operations
Sous Chef and Catapult announced a rollout covering more than 600 U.S. and Canadian restaurants. The focus is managing delivery operations. Deployment is ongoing; the vendor’s broader performance figures should not be presented as results already achieved by these locations.
A tile contractor buys robotic installation capacity
Foley Tile’s three-year, $3.55 million contract will deploy Tyler tile-setting robots while human installers retain cuts, finishing and quality control. The contractor pays by installed square foot instead of purchasing the machines. The precise rate is undisclosed, and productivity comparisons are the robot maker’s claims rather than independently audited results.
Free career guidance gets an expansion commitment
NextLadder’s $10 million commitment will support expansion of CareerVillage’s existing Coach service for young people. Individual access is already free. The announcement concerns funding over two years, not a brand-new application or a promise that using it will lead to a job.
|
|
|
THE ROUNDUP
Health and care
Microsoft’s medical AI moves toward hospital testing
Microsoft’s new Science publication and hospital partnership extend research on its diagnostic system to emergency cases and prospective evaluation. Earlier test results are retrospective; live hospital validation is still ahead. This is research progress, with no public patient app, price or firm clinical rollout date. It does not establish that the system outperforms doctors in routine live care.
A hospital’s AI pilot failed the full-workflow test
Nebraska Medicine’s account of a shelved cancer-registry pilot describes accuracy just above 80%, below its 95% target. A fallback summarizer also saved little time because nurses needed to check the records. These are a named executive’s reported results, not an independent audit. The lesson travels well: measure the checking and corrections alongside the generation time.
Tufts and an insurer target prior-authorization paperwork
Tufts Medicine and Blue Cross Massachusetts plan to use Humata Health for selected imaging requests, including MRIs. Their goal is to transmit required information in minutes rather than days. Faster paperwork does not guarantee faster approval, and the announcement is not independently measured evidence of reduced waiting times.
Rivanna announces clearance for ultrasound anatomy labeling
Rivanna says its BoneEnhance module received FDA 510(k) clearance for identifying bone and surrounding tissue with image overlays. That concerns anatomy labeling, not automated fracture diagnosis. Injury detection remains investigational, and the announcement does not supply evidence that the software improves patient outcomes.
Eko opens preorders for a lower-priced digital stethoscope
The CORE 300 product page shows a $199 promotional preorder, with estimated October 26–31 shipping. Optional AI features require Eko+, renewing at $119.99 a year after a 14-day trial. Eko’s AI software is intended for licensed healthcare professionals, not self-diagnosis. Buyers should consider the recurring cost as well as the device price.
AI requests for patient records add infrastructure costs
Epic says consumer chatbots’ requests for medical records are increasing health systems’ infrastructure costs. Its notification approach would let apps wait for a record change instead of repeatedly checking. The access discussed is patient-authorized; this is not a breach report, and savings have not yet been demonstrated.
|
|
|
THE ROUNDUP
Your accounts, software and buying decisions
Banks are seeing agents arrive without their blessing
Axios reports that banks are encountering third-party agents accessing customer accounts without the institutions’ approval. Customers authorizing an assistant and banks supporting that access are different questions. The reporting relies on unnamed sources and does not establish a tally of financial losses. Check your bank’s supported connections before delegating account tasks.
Railway investigates a possible exposure of application secrets
Railway’s security disclosure describes a potential exposure involving environment variables and secrets used by hosted applications. The scope remains under investigation. If your business runs software there, ask the person maintaining it to follow the provider’s precautionary credential-rotation and log-review instructions rather than waiting for confirmed exposure.
A compromised AI-software release puts credentials at risk
LiteLLM’s incident update describes a compromised release containing credential-stealing code. The investigation’s entry-point explanation is not yet a confirmed final finding. This matters most to organizations whose software uses the affected package: ask whether an affected release was ever installed or run, then follow the official response. Changing versions alone does not undo earlier credential exposure.
ANSI plans a register of evidence behind AI claims
ANSI’s announced AI Register would bring together vendor declarations, customer attestations and independent assessments. Eligibility details are still to come. A listing will not mean ANSI certified or endorsed the product, so buyers will need to distinguish which claims have independent support.
An outside review found gaps in internal agent safeguards
METR and Redwood Research published findings from a summer review of OpenAI’s internal agent monitoring. OpenAI says it patched findings affecting production safeguards; the reviewers have not independently verified those fixes. The study examined a historical system and does not establish a current ChatGPT breach.
Cyware turns threat reports into priorities for smaller teams
Cyware’s new agents rank vulnerabilities and prepare briefings for smaller organizations in participating threat-sharing networks. The sharing hub needs the Cyware platform; this is not a free standalone tool for any company. Automated remediation remains future work, and pricing is not disclosed.
|
|
|
THE ROUNDUP
Cars, culture and the world around AI
GM puts more AI into cars and vehicle development
GM announced an OnStar AI assistant planned for a future update to the 2027 Cadillac LYRIQ. Separately, the company says AI could help shorten vehicle development toward two years. Drivers should distinguish those future plans from features already in their cars; pricing and broad availability remain open questions. Development reporting.
Google adds AI help for advertising teams
Google’s October marketing update adds Gemini-assisted planning and advertising workflows. These are tools within its paid marketing platforms, not a free assistant automatically available to every small business. Check the particular feature’s rollout and platform requirements before assuming it is ready in your account.
Another listening pendant enters the market
Nirva plans November orders for a $249 listening pendant. Availability is still ahead; privacy promises deserve scrutiny.
Election questions need a source beyond the chatbot
New reporting examines misleading election answers and donation links in AI responses. It is a reason to check a specific answer, not a measured error rate for every chatbot. For registration, polling places and ballot rules, follow the answer back to the relevant election office before acting or sharing it.
Data-center developers face demands from their neighbors
AP reports on efforts to win community support through concessions and commitments around data-center projects. Residents have practical questions about taxes, costs and transparency. Read what a developer has committed to, who can enforce it and how results will be reported; a promise at announcement is not a delivered benefit.
Strikes on Russian data centers disrupt everyday services
AP’s new reporting describes how earlier October strikes on data centers in Russia disrupted shopping, ride-hailing and AI-related services. The fresh story concerns the impact, not a newly occurring attack. Avoid reading the account as proof that every named service remains offline. For businesses, it is a reminder to know what becomes unavailable when a shared provider fails.
A companion-robot maker responds to a privacy exposure
Following reporting about exposed media, Eiliko’s maker says it is reviewing security and disabling features. The response does not establish that every problem has been fixed. A friendly face on a household device does not answer questions about what it records or who can access it.
Mori takes a nonverbal approach to a home companion
Mori communicates through movement, light and sound rather than conversation. The lamp-like robot is a $649 preorder before tax and shipping, with late-December delivery estimated and two years of Nature Time included. Its wellness and privacy promises are company claims; this is not a product tested for this newsletter.
The Boston Symphony experiments with a live AI partner
Tod Machover’s “already and not yet” pairs human musicians with a system that responds to their playing. The October 22–24 premiere offers a human-led creative use of AI beyond generating a finished song. This account of the ongoing premiere run is not a performance review, and the whole work should not be described as AI-composed.
|
|
|
FOLLOW-UP
A new reason to check your website’s PHP update
Amazon’s new threat-intelligence account identifies attempts against PHP before the vulnerabilities were publicly disclosed. VulnCheck’s October 22 update also reports a successful controlled exploit under additional server conditions. The laboratory result and observed attempts do not establish successful criminal break-ins at customer sites.
If someone else maintains your website, ask whether its PHP installation is affected, whether the relevant fix or supported backport is installed, and whether the response calls for reviewing earlier activity. Use the official PHP release notices rather than assuming an automatically updated plugin also updates the server underneath it.
|
|
|
AI WORKFLOWS · FRIDAY: RAPID SKILL ACQUISITION
Learn to catch a misleading percentage claim before you repeat it
The new library-training initiative includes checking AI-generated claims. Here is one small skill you can practice now: telling a percentage change from a percentage-point change.
If customer satisfaction rises from 80% to 88%, the difference is eight percentage points. Relative to the starting 80%, it is a 10% increase. Both are valid, but they answer different questions. Getting the units right matters when you explain a dashboard or compare a promotion.
Allow 20 minutes. Use fictional numbers and a tutor you already have. ChatGPT Study Mode offers questions and guided practice across plans, subject to normal usage limits; Gemini Guided Learning is another option. No customer files or paid upgrade are needed.
Minutes 0–2: Try before being taught. Write how you would describe a response rate rising from 20% to 25%. Give both units and explain your calculation. Keep this first attempt.
Minutes 2–6: Correct the first mistake. Give the tutor your attempt and the prompt below. Percentage points come from subtracting the old rate from the new one. Relative change divides that difference by the old rate, then multiplies by 100. A zero starting value makes the usual relative calculation undefined.
Minutes 6–12: Practice one problem at a time. Ask for a rising rate, a falling rate and a changing count. Attempt each before seeing an answer. If stuck, take one hint and retry. Use a calculator to check the tutor’s arithmetic.
Minutes 12–16: Close the chat. Describe a completion rate rising from 40% to 50%, a return rate falling from 12% to 9%, and orders rising from 100 to 120. Give the units and show your work before checking the answers below.
Minutes 16–20: Check whether you can explain it. The completion rate rises 10 percentage points, or 25%. Returns fall three percentage points, or 25%. Orders rise 20%; they are counts, so percentage points are the wrong unit. If you missed one, practice that type again. Explain aloud why the starting value matters.
Copy this assignment: Help me practice percentage changes versus percentage-point changes. Start with my attempt below and identify the first mistaken step. Explain briefly, then give three fictional problems one at a time: a rising rate, a falling rate and a changing count. Wait for my answer. Offer one hint before revealing a correction. Show the arithmetic and units. Stop after those problems so I can take an independent check with the chat closed. My attempt: [paste answer and reasoning].
Only have five minutes? In an already signed-in chat, spend one minute answering this yourself: a rate rises from 10% to 12%. Allow two minutes for the tutor to check your reasoning. Close the chat and spend two minutes on a fresh 25%-to-30% example. Check afterward: the answers are two points/20% and five points/20%. Save your two worked answers as the result.
An AI explanation that sounds clear is not the finish line. For this narrow exercise, aim for correct independent answers and units. Try a fresh example tomorrow. The teaching sequence has not been tested with learners; 20 minutes is a practice budget, not a mastery guarantee.
|
|
|
TOOL SPOTLIGHT · EARLY LOOK
Dadan turns a walkthrough into instructions someone can follow
If you keep showing coworkers how to complete the same online task, Dadan’s guide creator is worth a small test. Record the process with its browser extension, and it captures clicks as numbered steps with screenshots and draft instructions. You can revise wording, annotate images, remove steps and share online or as a PDF.
It suits an owner, trainer or team member documenting a repeatable task, such as updating a product listing. Start with a harmless example, then ask a colleague to follow the guide without your help. Missing explanations will become easier to spot.
Cost and limits: free includes two guides a month, with up to five steps each. Lite lists $14 a month on monthly billing, with five guides of up to ten steps each. Longer procedures need a higher tier.
Skip it if your existing documentation tool already does this well. Dadan advertises automatic blurring of sensitive information; inspect every screenshot before sharing anyway. This is an early look at published features, not a hands-on review.
|
|
|
BEFORE YOU GO
One thing to try
Choose one recurring task before you choose another subscription. What would you need to see in the finished work to know the tool had actually helped?
|
|
mAIn Street · Aric Mitchell
Reporting checked October 22, 2026. Availability and prices can change. |
|
|