The Future Last Week: AI Can Do More Work. Who Gives It the Keys?

If you lead a team considering AI, the most important news this week was not a race for the highest model score. Meta released a stronger model built to pursue multi-step tasks. Workday announced agents that can keep HR and finance processes moving in the background. Google made the sources inside AI search answers easier to reach. In a laboratory, AI helped narrow a biological search, but the candidate still has to survive an experiment. Each story points to the same shift: the interesting question is less often “Can it answer?” and more often “What can we safely let it do, and how will we know it worked?”

This is an optimistic issue. I think useful AI is moving closer to the everyday work of companies, not farther away. But optimism earns its keep only when we separate a released product from a promised one, a plausible scientific lead from a validated result, and a power contract from electricity already on the grid. The reporting window is October 11–16, 2026; checked Saturday, October 17.

The week in one minute

• Meta released Muse Wonder and Muse Wonder Full on October 12. Wonder is available through its model interface for developers and Muse Code; Full is available to Muse Code subscribers. Wider Muse rollout is still ahead. The company says the models can remember relevant context and persist through a task; independent performance in your work remains to be shown.
• Workday announced background agents for hiring, payroll, accounting, and compliance on October 13. It says customers choose the approval points and can review action logs. Ask which functions are actually available in your contract and what an agent may change without review.
• Google updated AI Mode and AI Overviews on October 15 to put more source links inside answers and higher on the page. Some related features are tests or rolling out by location. A link being present is not the same as a publisher receiving a visit.
• Google DeepMind described an AI-assisted antimicrobial search on October 15. The identified molecule is a candidate, not a medicine; laboratory validation is underway. Separately, the Simons Foundation announced a $15 million joint effort with Anthropic and OpenAI to help mathematicians evaluate AI-assisted work.
• The Associated Press examined three proposed U.S. nuclear plant restarts tied in part to rising data-center demand. Together, they represent a planned 2.25 gigawatts, not power currently produced by restarted units. Licensing, safety, cost allocation, and dates still matter.

1. Meta offers a longer-running model. A benchmark is not a business process.

Meta says Muse Wonder and Muse Wonder Full are designed for tasks such as understanding a lease, navigating an insurance claim, and building an application. Two product ideas are more interesting than the launch adjectives: “Active Memory,” which carries relevant information across conversations, and “Active Focus,” which keeps trying when a task stalls and asks before consequential actions. Those are Meta’s descriptions of how the product is intended to behave, not a guarantee that it will handle a company’s exception queue correctly.

The availability distinction matters. Muse Wonder can be tried by developers through the Meta Model API and Muse Code now; the Full version is limited to Muse Code subscribers. Meta says more consumer uses will follow. Axios noted that the next test is whether Meta’s reported performance holds up in independent evaluation. The company also says a fuller safety report is still to come. If a sales deck collapses those caveats into “available to everyone and best at everything,” you have learned something about the sales deck.

What is real, what is hype, and does it apply to you?

Real: another major supplier is offering a model designed to continue work rather than answer one prompt. That gives buyers more options and may pressure prices and product design. Hype: treating vendor comparisons or a demonstration as proof that an agent can safely negotiate a claim or interpret a contract on its own. If your company builds internal tools, handles long document reviews, or serves customers with multi-step requests, this is worth a controlled trial. If your main need is drafting a weekly summary, there is no urgency to switch. Give the new model twenty completed cases and twenty troublesome ones; compare accepted answers, corrections, elapsed time, and the full cost of review against your current system.

Watch for an independent test on work resembling yours and the promised safety report. The stronger signal is not a leaderboard position. It is a lower rate of consequential errors on cases your people once had to rescue.

2. Workday moves the agent into a process that already has an owner

Workday’s October 13 announcement gives a more concrete view of “AI at work.” A recruiting agent can watch open jobs, flag a thin candidate pool, and draft outreach, but the recruiter approves the message before it goes out. Workday says customers can choose what runs automatically, what requires review, and how each action is logged. In payroll and accounting, that distinction is not a nicety. A wrong recommendation can be corrected; a wrong payment or employee record creates a different problem.

Workday also introduced Work Intelligence, an early-access tool for breaking jobs into tasks and planning where people should be retrained, and announced an agreement to acquire C1.ai, which manages access for people and AI agents. The acquisition has not closed. These three moves belong together: an agent needs an assignment, a defined permission set, and a person who remains responsible for the outcome.

What is real, what is hype, and does it apply to you?

Real: Workday is putting persistent assistance inside the records and approval routes companies already use. Hype: assuming “autopilot” means zero supervision, or that every announced feature is live for every customer. Workday executive Gerrit Kazmaier put the tension plainly: “in payroll and finance, a good guess isn’t good enough.” That is a vendor voice, but it states the buyer’s problem accurately. This applies most directly if HR, finance, or IT already runs on Workday. Others can use the same design test with a different supplier: what records can the agent read, what can it change, who approves the risky step, and can a reviewer reconstruct what happened?

A small Workday user discussion this week mentioned a practical obstacle: useful knowledge can sit outside Workday and remain inaccessible to the agent. That is firsthand feedback, not a representative survey. It is still a good pilot question. Test one real task that crosses a system boundary; count manual handoffs and failed permissions. Watch for documented customer results, not just a launch-stage workflow diagram.

3. Google gives sources more room in AI Search. Will readers use the door?

Google says it is adding more links directly inside AI-generated search answers and placing them higher more often. It will test a prominent section for official sites, expand timely article carousels to additional places and languages, and make related-article links more visible. These are product changes, not evidence that referral traffic has recovered. Google says use of its Preferred Sources feature grew more than sevenfold since the start of 2026, but it did not publish the starting number in this announcement. The ratio does not tell a publisher how many visits arrived.

For a business reader, easier access to source pages is welcome. If an AI answer summarizes a rule, a product limit, or a medical claim, the original page is where the date and conditions live. For a publisher, this is an opening, but not a rescue plan. The skeptical view is fair: more visible citations can coexist with fewer visits if most people stop at the summary. Independent research on AI search and publisher referrals argues that reduced referral traffic is a measurable concern, not merely a publisher mood.

What is real, what is hype, and does it apply to you?

Real: Google changed link placement and presentation in AI search. Hype: claiming this will automatically restore organic clicks. If your company publishes useful research, support material, or product documentation, make the original page worth the next click: show the method, date, examples, downloadable tool, or decision table that a summary cannot replace. Track search impressions and visits by page, then look for assisted inquiries or qualified readers. Watch whether prominent source links actually increase visits for your pages over several weeks, rather than judging a rollout from a screenshot.

4. Science shows both the promise and the missing last step

Google DeepMind described a scientist directing an AI workbench across protein databases, structural evidence, and published research to find a possible antimicrobial peptide—a small protein that might act against bacteria. The agent helped narrow an enormous search to a candidate and draft an experimental plan. The essential sentence in the company’s own account is that the lead is now undergoing wet-lab validation. No treatment has been demonstrated. No patient outcome follows from a promising computer prediction.

The useful pattern is broader than drug discovery: AI can compress the hunt for promising possibilities while a domain expert decides what evidence is sufficient and what physical test comes next. A manufacturer might use the same shape of work to rank suspected failure causes before testing parts. A financial team might prioritize unusual transactions before an accountant reviews the source records. Those are analogies, not outcomes shown by the DeepMind experiment. The common failure is to promote a ranked lead into a confirmed result without the expensive verification step.

The scientific community is building that step into the system. On October 15, the Simons Foundation announced a combined $15 million commitment with Anthropic and OpenAI to support mathematicians evaluating AI-assisted results, learning tools, and shaping standards. Foundation president David Spergel said the community, “not any single company or institution,” should drive those discussions. That is a constructive outside reaction to a month of ambitious AI-math claims: fund understanding and review, not just more output.

What is real, what is hype, and does it apply to you?

Real: AI can help a skilled researcher search and connect evidence faster; the foundation is committing resources to independent evaluation. Hype: confusing a candidate with a cure, or an AI-written proof with mathematical understanding. Most executives do not need a science model next Monday. They do need the habit illustrated here: name the verification step before scaling the discovery step. Watch for laboratory results on the peptide and publicly reviewed mathematical work that other experts can reproduce.

5. The power behind the screen is getting a line item

Associated Press reporting on October 16 brought three shuttered U.S. nuclear plants into one picture: Palisades in Michigan, Crane in Pennsylvania, and Duane Arnold in Iowa. The U.S. Department of Energy has committed loans of up to $4.4 billion toward the restarts, according to AP, with a combined planned capacity of about 2.25 gigawatts. Two projects have long-term technology-company customers. None of that means all three are generating again today. Duane Arnold, for instance, is being targeted for early 2029 in the report.

This is where an “AI infrastructure” chart meets a town’s budget, a regulator’s permit, and the weather. University of Michigan nuclear engineer Todd Allen told AP that existing, previously connected plants can return power more quickly than entirely new reactors. Edwin Lyman of the Union of Concerned Scientists told AP he worries safety upgrades are not receiving enough attention. Those are substantive competing judgments. AP also reported a split over public financing: supporters see leverage for needed supply, critics question using taxpayer money where private demand is strong.

What is real, what is hype, and does it apply to you?

Real: companies and governments are financing concrete electricity projects in response to demand, and some long-term contracts exist. Hype: adding planned megawatts as if they are connected, licensed, and affordable power now. You do not need to be a utility executive for this to matter. AI service cost, regional capacity, delivery dates, and reliability can affect contracts and operations. If you are buying a large AI service or planning a data-heavy facility, ask the supplier where capacity comes from and what happens if delivery slips. Watch permits, grid connections, actual generating dates, and who pays for upgrades.

A quieter policy signal: AI will be asked about elections

Anthropic examined a random sample of 250,000 U.S. Claude.ai consumer conversations from September 21–27 and published the analysis October 14. It classified about 5.75% as touching broadly on politics, public policy, or social issues; about 0.51% were specifically about elections or voting. This measures one provider’s consumer conversations over one week, not the public’s political beliefs or all AI use. Anthropic says its existing voter-information banner links to TurboVote, a Democracy Works service; it plans to add state-specific voting information directly in answers later in October. It also says it banned accounts involved in deceptive political content. Those safeguards are company-reported actions, not a claim that election misuse is solved.

For businesses, the lesson is not that every conversation is dangerous. It is that assistants are becoming a normal place to seek consequential information. If your organization serves the public, test whether your AI answers give dates and direct source links for changing, high-stakes facts. Do not let a polished response outrank an official record.

What people are saying—and what those voices can tell us

This week’s reactions are useful precisely because they do not agree. Meta’s release says its models can persist through difficult tasks; Axios asks whether the performance survives independent testing. Workday’s product leader says a guess is insufficient for payroll. Spergel at the Simons Foundation argues researchers should shape evaluation, rather than leave it to a vendor. Allen and Lyman, speaking to AP, see different risks in nuclear restarts. A handful of early forum comments about Workday agents raise integration questions, but cannot establish how customers as a group feel. These voices do not cancel one another out; they tell us where to put the test.

Five useful moves for Monday

1. Choose one multi-step task with a real owner. Write its starting input, permitted records, permitted actions, approval point, and finished outcome on one page before testing a new agent.
2. Run the same twenty ordinary cases and twenty failure-prone cases through your current approach and one new model. Have the process owner count corrected outputs, reviewer minutes, and complete cost. Pause if the new system creates an unreviewable action.
3. If a vendor says an agent works inside HR or finance, ask for a live demonstration using your exception case—not the polished default path. Confirm which features are live in your contract and export one complete action log.
4. If search matters to your business, improve one important source page so a reader can verify a claim and take a next step there. Record its impressions, visits, and qualified inquiries before and after Google’s rollout; do not mistake visibility for a click.
5. For one AI-dependent expansion, write down the physical or organizational bottleneck: electricity, system access, source data, specialist review, or training. Name the person who owns it and the evidence that would show it has cleared.

If you need a starting point for that first test, use our AI business-case worksheet to state the current cost and the outcome you would count. For the permission and approval design, the AI security course walks through the controls without requiring a technical background.

My bet is that the next competitive advantage will come from a company that makes agents boring in the best sense: they know their job, use only the keys they need, show their work, and stop at the right moment. This week brought more capable engines. The companies that build the checks around them will get more of the value.

Source and method notes

Checked October 17, 2026. Product access, planned features, safety descriptions, and model performance are attributed to the vendors’ releases. The DeepMind peptide is a computational candidate awaiting laboratory validation. The Simons Foundation’s $15 million is a combined commitment of funding and resources, not a grant already spent. AP’s nuclear figures describe proposed restarts and loans, not completed electricity production. Anthropic’s conversation analysis covers its U.S. consumer app in a sampled week; it is not a population poll. External commentary and firsthand forum feedback are identified where used. The hero is an AI-generated editorial metaphor, not documentary evidence. Links appear alongside the claims they support.