25 AI developments, a lesson on defining your first product, and a tool for predicting which parts of an ad draw attention.
 |
| Friday, October 9, 2026 · Edition 329 |
Google is bringing one agent across workplace apps, Amazon is putting Alexa alongside what you see on a tablet, and OnDeck is extending automated loan decisions to new customers. Also inside: proposed AI safeguards, 20 more developments and Friday’s Claude Startups workflow. |
|
TOP 5
Five stories worth your time.
|
|
01
Google announced a Gemini agent designed to handle work across Gmail, documents and connected business apps. According to Google, it can research a question, build a spreadsheet or presentation, and keep working in the cloud after you close your laptop. Teams can also give an agent its own role and limit the information it can access.
The system can choose between Google and Anthropic models, while project spending caps pause work when the budget is reached. The announcement did not provide a universal rollout date or plan-by-plan pricing. For businesses, the useful question is whether those permissions and cost controls fit the work they would actually delegate.
Google / official announcement · Oct. 8
|
|
|
02
Amazon’s new Alexa Tablets put its AI assistant alongside whatever you’re reading, watching or working on. With screen access enabled, Alexa can use the page in front of you to answer questions, pull dates into a calendar or help find a product. Screen sharing is off by default and has a visible indicator when active.
The Android tablets include the Google Play Store. U.S. shipping is scheduled to begin October 14. Starting prices are $229.99 for the smaller model, $329.99 for the 11-inch version and $499.99 for the 12 Pro. The keyboard and pen cost extra. These are announced capabilities; independent testing will show how reliably they work.
Amazon / launch announcement · Oct. 8
|
|
|
03
OnDeck is extending automated final loan decisions to new small-business applicants after rolling the process out to eligible existing customers. The lender says its AI verification and machine-learning systems can reduce decision time from a few hours to under a minute for qualifying applicants. Some approved borrowers can proceed to funding without supplying additional documents.
The company says its existing underwriting standards still apply and that it monitors the models’ performance. The announcement concerns the application process, with no guarantee that every borrower will qualify or receive money within a minute. It also provides no evidence that faster decisions produce cheaper loans or better borrowing terms.
OnDeck / company announcement · Oct. 8
|
|
|
04
Sen. Maria Cantwell, the Senate Commerce Committee’s top Democrat, has outlined a six-part framework for regulating advanced AI. It calls for federal safety standards developed by NIST, independent audits before covered models are released, continuing tests and prompt reporting of serious failures.
For workers and consumers, the proposal calls for meaningful human oversight when AI influences employment, health care, credit or other essential services. People would be able to understand how AI affected a decision and challenge it before a qualified human reviewer. This is a policy framework, with no legislative text or new legal requirements. Turning those principles into enforceable protections would require further action.
Senate Commerce Committee / Sen. Maria Cantwell · Oct. 7
|
|
|
05
HP is testing a version of its website that rearranges products and comparisons as shoppers explain what they need by voice or text. Someone shopping for a travel laptop can describe their priorities, answer follow-up questions and see the options change. A business buyer can work through requirements such as security, battery life and video calls for a hybrid team.
HP Experience AI is in beta with a limited group of U.S. customers. Visitors can join the early-access waitlist from the bottom of HP.com. A wider global rollout is planned later this year. HP has not published results showing whether the new experience improves shoppers’ decisions.
HP / official launch post · Oct. 8
|
|
|
AI WORKFLOWS | FRIDAY: RAPID SKILL ACQUISITION
You can learn to define an AI product before spending your startup credits.
On October 6, Anthropic expanded Claude Startups, giving approved founders more help building a company on Claude. The offer includes $1,000 in API credits, a free year of Claude Team and access to Applied AI office hours.
For a nontechnical founder, one useful starting skill is product specification: explaining what a tool must do clearly enough that someone can build it and check its work.
Suppose you run training workshops. You want a small product that turns workshop notes into a follow-up task list. “Make a useful summary” leaves important decisions to the AI. Your first lesson is writing rules you can actually test.
| THE PRODUCT YOU ARE LEARNING TO SPECIFY |
| Workshop notes |
→ |
Draft task list |
→ |
Trainer approval |
| Every task points to its source. Missing owners stay unassigned. Conflicting dates get flagged. |
On a paid Claude plan, use approved, read-only connections to retrieve examples and create a private artifact. Company settings may restrict these tools. If business material is off-limits, have Claude invent fictional workshop notes. You supply the goal and decide what counts as correct; Claude organizes the examples, builds the demonstration and prepares the checks. Keep the demonstration disconnected from live business systems.
-
Try the skill first. Write three pass/fail rules before asking for suggestions. Include what happens when information is missing.
-
Learn through a small build. Have Claude turn your rules into a one-page product brief and private demonstration. Ask it to explain each proposed change to your rules. Approve the rules before it judges the output.
-
Examine the failures. Have it produce a test sheet covering clear notes, missing owners and contradictory deadlines, with expected and actual results. Inspect the source passages yourself. Broader testing is needed before customer use.
-
Prove the learning. Close the coaching conversation. Write a brief and three testable rules for a different task, such as extracting unanswered questions from workshop feedback. Ask a colleague to check whether the rules resolve ambiguity without inventing information.
Give Claude this assignment
Help me learn to specify a small AI product for workshop follow-up. Use only approved sources through read-only tools, or fictional material if access is unavailable. Review my initial rules, explain the gaps, then build a private brief, demonstration and test sheet. Separate expected results from observed results. Tell me which checks you could not run. Exclude customer messages, publishing, spending and live record changes.
Check the offer before applying
Bootstrapped startups can apply if founded within five years or funded within two. You need a website, matching company email, Claude Console account and description of what you are building. Anthropic reviews applications; an idea alone does not establish eligibility.
The Team offer covers up to five Premium seats for organizations new to Team. API credits expire six months after being granted. They cover direct Claude API usage in software you build. Partner offers advertised at up to $45,000 have separate terms and no cash value. Review the program terms before applying through Claude Console.
For owners, have a colleague review the brief before assigning implementation. Accepted members can take one unresolved design question to office hours; availability varies.
Choose one task you know well and write its first three pass/fail rules today.
|
|
|
AI TOOL SPOTLIGHT · TractionWave
TractionWave adds a second opinion before you publish an ad, flyer or social post. Upload a PNG or JPG and it returns an AI-predicted attention heatmap, plus a ranked list of areas likely to attract the eye. A shop owner could use it to check whether a sale price or booking button gets buried under a large photo.
Treat the result as a design clue. It predicts attention; it doesn't measure actual customers or tell you whether they will buy. Compare its suggestions with real campaign results.
One watermarked evaluation is free without an account. Signing up adds one watermark-free evaluation; paid packs start at $18 for 10 credits, with no subscription. Start with the web version. Its Canva integration is still marked coming soon.
TractionWave / product and pricing · Oct. 8
|
|
|
MORE AI NEWS
Headlines
|
Customer service and operations
Sanas introduced a customer-service system in which one employee supervises up to four AI phone conversations. The person can listen, guide the AI privately or take over mid-call. Customers are told AI is involved, and actions are recorded. Access is by invitation now, with a wider release planned for December.
Sanas · Oct. 8
Cisco announced AI teammates that can join Webex conversations and carry out work across approved business apps. Claude support is coming, while a new customer-service system called Dialog is designed to keep working after a call ends. The announcements include access controls and activity monitoring; several features have future release timing.
Cisco · Oct. 7
Sierra’s new fleming-1 model flags likely synthetic voices during customer-service calls. A synthetic voice is not itself evidence of fraud. The company is also developing a standard with Meta and partners including Shopify, Stripe and Walmart to identify whom a personal agent represents and what it may do. The first specification is due later this month.
Sierra / product announcements · Proposed agent standard · Oct. 8
Ask TRR is expanding beyond its pilot to all 45 million RealReal members. The Gemini-powered assistant lets shoppers describe an occasion, style or other preference without knowing the right designer name or search term. It searches the reseller’s changing inventory of unique items. AI recommendations on individual product pages remain a planned addition.
The RealReal and Google Cloud / launch announcement · Oct. 8
Infor’s October release adds specialized agents across its business software. For smaller distributors, examples include preparing quotes, tracking late purchase orders and checking stock. The release also expands AI governance and access to product documentation. Infor says the agents work within approval processes, with capabilities varying by the company’s specific software package.
Infor · Oct. 6
|
|
Business data and controls
New Zealand’s LawVu launched Lens as an add-on for LegalOS customers. Legal teams can ask questions across their contract collection and trace answers to the underlying agreements. The company points to renewal terms, service commitments and acquisition obligations as uses. Those findings still need legal review; the announcement provides no independent accuracy test.
LawVu · Oct. 6
Rencore opened a preview that connects Claude Enterprise oversight with Microsoft 365 and Copilot management. IT teams can find unused accounts, publicly shared projects and spending by employee or department. ChatGPT Enterprise support is planned later this month. The service tracks metadata and account activity, rather than reading employees’ messages or files.
Rencore · Oct. 7
Apollo introduced GraphOS Agent Services to control which business data an AI agent can see and which actions it can take. Rules can apply to individual fields, with activity logged for review. Intuit is piloting the service in preview. The approach puts access decisions in explicit software rules instead of leaving them to a model.
Apollo GraphQL · Oct. 7
Dell introduced additions to its AI Data Platform that connect company records with the meaning behind business terms. Its example is inventory: a part on a shelf may still be unavailable if it is already allocated or on a quality hold. An accompanying agent can show supporting evidence and seek approval before acting.
Dell · Oct. 6
CAEVES announced Context LAKE, which indexes business files where they already sit and passes relevant information to connected AI assistants. The company says existing file permissions remain in force and answers include their sources. Release is scheduled for October 20, including deployments in a customer’s own data center or private cloud.
CAEVES / product announcement · Oct. 7
|
|
Building and running AI
Harness acquired selected Augment Code assets, including Cosmos and its code-context technology. The resulting Cosmos agent can turn a work ticket into proposed code, tests and a review request, then respond to feedback. Teams set approval boundaries. Harness says Cosmos is available now; the announcement covers selected assets, not the entire company.
Harness · Oct. 8
Perplexity released two downloadable models for finding information in text, images and document pages. A smaller model can search an index built by the larger one, potentially reducing the computing needed for each query. They are building blocks for search applications, with company-reported benchmark results, rather than a new feature automatically appearing in Perplexity chat.
Perplexity / model card · Oct. 7
Architect’s Liquid Inference lets computing providers compete to answer AI requests. Buyers can set cost limits, response-time requirements and rules about where data is handled. The platform then selects an eligible offer and locks a maximum price before work starts. It is a new way to buy model access, without a demonstrated savings guarantee.
Architect · Oct. 7
Silicon Labs opened a beta that gives coding assistants access to its device documentation, development tools and connected hardware. Initial support centers on Bluetooth devices. It also connected edge-AI tools with Databricks so teams can manage device data and model development together. A separate hardware-design assistant is planned for an alpha release in January.
Silicon Labs · Oct. 7
|
|
Work and public life
In Indeed’s U.S. survey, 74% of employers called effective AI use a fundamental skill, but only 27% had a consistent written definition across roles. Nearly a third of job seekers reported unexplained AI requirements in listings. The study covered 500 hiring decision-makers at larger companies and 1,007 job seekers, with fieldwork in July and August.
Indeed / YouGov research · Oct. 6
Michael Smith was sentenced to 18 months in prison after pleading guilty to a music-streaming fraud conspiracy. Prosecutors said he used AI-generated songs and automated listening accounts to collect more than $8 million in royalties. The court also ordered roughly $8.09 million in forfeiture and two years of supervised release.
U.S. Department of Justice / sentencing announcement · Oct. 6
Oakland’s City Council approved a 45-day pause on new data-center approvals while the city studies rules for energy use, water use and neighborhood impacts. The temporary moratorium took effect immediately and can be extended. It gives local officials time to examine development pressures associated with the growth of AI computing.
City of Oakland / council record and ordinance · Ordinance text · Oct. 6
|
|
Science and health research
The White House announced $2.4 billion in industry commitments of AI tools and computing credits for the Genesis Mission Consortium, supporting research across more than 15 federal agencies. Contributors include NVIDIA, AMD, OpenAI, Anthropic and Google. The wider announcement also included regional computing initiatives. These are announced commitments, with benefits depending on delivery and researchers’ access.
White House / science summit fact sheet · Oct. 8
Biohub’s Virtual Biology Initiative now combines $1.8 billion in funding, data, computing and measurement resources. New participants include DOE, NIH, Google DeepMind, Isomorphic Labs and Meta. The aim is open data that can help models predict how cells respond to interventions. The total includes earlier commitments and existing research assets; it is not all new cash.
Biohub / partnership announcement · Oct. 7
SpliceAI2 predicts how genetic changes affect the way cells assemble RNA. Illumina says it identified 17% more disease-relevant splice variants than comparison models in a rare-disease research dataset. The model is available through its BioInsight applications. The evidence comes from the company’s newly released preprint and does not establish improved patient outcomes.
Illumina / model announcement and preprint · Research preprint · Oct. 8
|
|
|
mAIn Street gives nontechnical readers useful AI news they can put to work.
mAIn Street #329 · Friday, October 9, 2026
Reporting checked through October 8, 2026.
|
|
|
|