The Future Last Week: AI Agents Get Jobs, and the Work Gets Real

AI agents that keep working after you close the chat window became a serious product category this week. OpenAI introduced Dots; Meta brought Muse into small-business tools. Google showed a powerful new model but held back broad access, and Anthropic made a faster model available now. The question for a business leader is shifting from whether an assistant can produce a good answer to whether it can finish an approved job, with the right information and a clear point for human review.

This second issue of The Future Last Week covers announcements from September 28 through October 2. The mood is hopeful because the choices are improving. The limits are just as concrete: product access varies, company examples do not prove a result will transfer to your team, and agents still need trained owners, electricity, and local trust.

The week in one minute

• OpenAI introduced Dots, persistent agents with connected apps, at its September 29 developer event. Meta expanded Muse for small businesses the same day. Both products promise work that continues across tools; neither announcement establishes the cost or reliability of your finished task.
• Anthropic released Claude Sonnet 5.5, which it says is faster and cheaper to run than Sonnet 5 for typical work. Google introduced Gemini 4 Argon with access limited first to trusted cyber defenders. Test availability before planning a rollout.
• Barclays and Anthropic described an internal knowledge assistant used by more than 16,000 colleagues and an email process handling about 120,000 messages daily. These are company-reported activity counts, not measured customer savings.
• Anthropic committed $100 million to a training program aiming for 10,000 engineers by the end of 2027. The first cohorts are running; the target is not a completed result.
• Amazon said it will add more than $1 billion over five years to programs in U.S. communities hosting its data centers. It is a commitment and a response to physical and local constraints, not proof that every project has community approval.

1. Persistent agents make the permission decision unavoidable

OpenAI says its new Dots can take ongoing responsibilities, use connected apps, and keep working on a person’s behalf. Its DevDay recap says Dots are available on Pro and Business Premium in eligible markets; Enterprise, Education, and Healthcare workspaces can try a beta if an administrator enables it. OpenAI also announced a software interface for developers to build agents that operate other applications through their visible controls. Those are useful capabilities, but the recap is a product announcement, not an independent study of completed business work.

Meta’s Muse for Small Business announcement adds connections to business tools including QuickBooks, Shopify, Slack, and Stripe. Meta says Muse can bring together sales, campaigns, and accounts and draft work proactively. It says nothing publishes, sends, or spends without approval. That is the right boundary to describe clearly. A draft growth plan and a paid ad are different acts; a connected accounting system may expose information a marketing request does not need.

Does this apply to you?

Yes if a team already repeats a task across several approved systems. Pick one job, such as preparing a weekly customer follow-up list. Name the source records, the person who accepts the list, and any action that must wait for approval. Run it on real but low-risk cases first. If there is no clear owner or the source data is unreliable, a persistent agent will keep producing work that someone must untangle.

Hype versus real

Real: two large providers now put continued, connected work at the center of their products. Hype: “always on” means “always useful.” A task that runs without waiting can consume more tool calls, create more review work, or act on stale context. Measure the accepted outcome and the corrections, not how long the agent stayed busy. Our business-case worksheet separates released staff capacity from actual cash savings.

2. The model choice widened, but the release dates matter

Anthropic released Claude Sonnet 5.5 on September 28. It lists a price of $2 per million input tokens and $10 per million output tokens, and estimates that a typical token-billed workload costs up to 30% less than Sonnet 5. It also says the model runs 30% faster. These are Anthropic’s claims and pricing terms as of October 3, not a prediction of your total bill. A different number of attempts or more reviewer time could erase a token saving.

Google announced Gemini 4 Argon on September 30 for complex coding, professional work, and cyber defense. Its first rollout is to selected cyber defenders; Google says wider developer, enterprise, and consumer access will follow. Google reports strong benchmark results and internal examples, including memory optimization and code migration, but some of those projects still require manual audits before production. A price announced for a future broad launch is not the same thing as a model your team can buy and test today.

OpenAI also introduced GPT-6.1 Sol, describing performance near its more expensive Astra model at one-fifth of Astra’s standard input and output token prices. Three vendors thus made different offers in one week: an available everyday model, a restricted frontier preview, and another price-performance option. The useful comparison is not their launch charts. Replay the same work with the same acceptance rule, then count the total model charge, elapsed time, and human repair.

Does this apply to you?

If you already pay for repeatable AI work, set aside a small sample of real cases and re-run it on an available alternative. Include easy and hard cases, and let the same reviewer judge both without knowing the model. If you have no stable task or review rule, hold off on a model switch. Our model retest guide explains how to make the comparison useful.

Hype versus real

Real: model quality is rising at several price points, and buyers have more room to choose. Hype: a long output limit, a top benchmark rank, or an introductory price guarantees a cheaper completed report. Google’s phased release is also a reminder that an announced capability and general availability are separate facts.

3. Barclays shows the value and the missing measurements

The Barclays–Anthropic announcement offers a rarer look at work already under way. The companies say more than 16,000 Barclays UK colleagues have used a knowledge assistant that has handled more than one million searches since going live in 2025. In global markets, a system classifies and routes about 120,000 incoming emails each day. Barclays expects Claude Code to reach half of its developers by the end of 2026. The first two are reported activity measures; the third is a target.

Those counts matter because they describe repeated work at scale inside a regulated bank. They do not tell us the share of answers accepted, the change in customer resolution time, the error rate, or the cost of review. In the company announcement, Barclays executive Anne Marie Darling said the measure is the effect on “customers, clients, and colleagues.” That is the right test, but the announcement does not provide the before-and-after numbers to complete it. A knowledge search that saves an employee a step is promising; a customer issue resolved accurately is the outcome to verify.

Does this apply to you?

Yes if your people spend time finding policy answers or routing incoming requests. Start with one queue and record the current response time, correct-route rate, escalations, and customer recontacts. After a trial, compare the same measures on similar cases. Stop expanding if the new process makes staff faster but causes more wrong answers or repeat contacts.

4. The talent story is about shipping, not collecting badges

On October 2, Anthropic announced Claude Frontier Academy, backed by a $100 million commitment and a goal to train 10,000 engineers by the end of 2027. Its first cohorts include engineers from consulting firms and large companies. The program starts with a simulated deployment and a practical assessment, then a 12-week residency around a real project at the engineer’s own organization. Anthropic says the first advanced badges are expected in early 2027; the goal and credentials are prospective.

The interesting part is the assigned project. A trained engineer who returns to an organization without access to the right systems, a security owner, or a business sponsor cannot turn a course into a live service. The company has an incentive to grow Claude use, so this is not independent proof that its training pays off. Still, the design points to a real bottleneck: someone has to join process knowledge, software, testing, and adoption. Our analysis of hidden pilot work describes the work that sits between a demo and a dependable process.

Does this apply to you?

If you fund AI training, require each participant to bring back one narrowly defined project with a sponsor, baseline measure, and review date. Judge the result after it is used in ordinary work. A badge may identify a person to call; it does not replace a production check.

5. The data center bargain is now part of the AI product

Amazon said on October 2 that its Built Together program will add more than $1 billion over five years to education, job training, energy affordability, water preservation, and locally chosen priorities in U.S. communities hosting its data centers. The number is a future company commitment, not an audited sum already delivered. It follows last week’s reminders that AI capacity depends on grids, construction, and the people living near new facilities.

There is a constructive reading: the company is putting money and public terms around effects communities have asked it to address. There is also a skeptical reading: a national fund does not answer what one town will pay for a substation, how much water a particular campus will use, or whether new work will go to local residents. Those questions need project-level answers. No amount of model progress can bypass a delayed connection to the grid.

Does this apply to you?

Most AI buyers do not need to negotiate a data-center permit. They do need to test whether a provider can meet a promised price and service level as demand grows. Facility builders and local officials should ask for a dated power and water plan, the party paying for upgrades, and a way to check delivery against the community commitments. The relevant unit is the proposed site, not a nationwide pledge.

One smaller signal to watch

A September 28 paper led by Alan Chan and Christoph Winter, with researchers from several AI labs and universities, asks what could happen if AI increasingly automates the research that improves AI itself. The authors see a possible feedback loop that could speed progress and risk, but say the evidence remains uncertain. They also describe physical experiments and supply chains as possible brakes. This is a scenario and policy argument, not a report that an “intelligence explosion” has happened. For most business leaders, the near-term lesson is narrower: track which tasks your AI systems can actually complete and where independent checks still need people and time.

What people are saying—and what it tells us

Tom Mulholland, a grocery owner quoted by Meta, wants more time for the work that makes his business money. Barclays’ operations leader asked for effects on customers and staff. Researchers from several labs and universities warned in their September 28 paper that faster AI research could also compress the time available to check new systems. These voices speak to different scales of the same issue: speed helps only when people can verify what changed. The Meta account is promotional customer testimony, the Barclays figures are company-reported activity, and the research paper is a conditional argument. None alone proves a general return or a prediction.

Five moves for Monday

1. Choose one ongoing task for an agent. Name its approved sources, the person responsible, and the exact actions that require a human sign-off.
2. Compare an available new model with your current one on the same real cases. Count accepted outputs, review minutes, failed attempts, and the complete bill.
3. For a knowledge or routing assistant, measure correct answers, escalations, and customer recontacts before and after the pilot.
4. Tie AI training to one sponsored project and a review date. Ask for a working process rather than a certificate count.
5. If infrastructure affects your plan, ask for site-level power, water, cost, and delivery commitments rather than a provider-wide headline.

The signal to watch next week is whether persistent agents complete ordinary tasks with fewer corrections and clear permission records. That is a more useful sign of progress than the number of hours they can stay awake.

Source notes

This issue uses announcements published September 28–October 2, 2026. Product performance and pricing are attributed to providers and may change. Barclays activity counts and the small-business quote come from company announcements; neither is an independent impact study. The training and community-spending figures are commitments. Each changing fact links to its original announcement or paper. The illustrations are interpretive artwork, not records of real events.