The year agents went to work.

Washington showed what being a maker means. In June, US export controls forced Anthropic to suspend access to its newest models, Fable 5 and Mythos 5, until the Commerce Department lifted the controls on 30 June. In October the FBI arrested an OpenAI engineer after staff found a listening device in an executive meeting room at the company’s San Francisco offices, which puts espionage firmly in the contest. Britain’s Tech Prosperity Deal stayed on hold over trade disputes, with only its nuclear strand restarted.

The loop became the year’s defining fight between a lab and a state. The Pentagon designated Anthropic a supply-chain risk after the company asked for assurances that its models would not run fully autonomous weapons or mass domestic surveillance. At the same time the Pentagon chose Shield AI’s Hivemind for its low-cost attack drones, software described as letting swarms coordinate and decide without human intervention, and Baykar showed swarms finding and striking targets with GPS jammed.

The frontier is already rebuilding. By August, agents directed by Anthropic staff were logging six times their handlers’ estimated working hours, and the money is following, with AI-native software spending up 94% year on year in the first quarter against 8% for traditional SaaS. Most organisations are still stuck between pilot and production, so this holds for the leaders first.

Investors moved first. About $285bn came off software, financial services and asset management stocks in a single day in February after Anthropic released a new automation tool, and a survey of software companies found the share charging purely per seat had fallen from 21% to 15% in a year. Salesforce, down about a third by May, is repositioning around Agentforce. Per-seat pricing does not survive when agents do the work.

Anthropic has now shown it in its own numbers. Its employment study, published on 16 October, shows the junior share of technical hires falling from 27% to 9% since early 2025 and the junior share of technical staff from 21% to 14%, while technical workers spend more of their time managing agents and in meetings. That growing layer of people directing agents is the diamond’s middle. Outside research is less clear that AI alone is cutting junior demand, so read this as the frontier moving first.

Data centres turned AI into a voting issue in the US midterms. Kathy Hochul imposed the first statewide moratorium on new hyperscale data centres, and Greg Abbott moved to restrict them in rural Texas. Bernie Sanders, Steve Bannon and Josh Hawley are all warning that AI threatens working people, which is the coalition of unlikely allies the report expected.

The money came first. Yann LeCun’s AMI Labs raised $1.03bn in March, Europe’s largest seed round, before shipping anything. Commercial use is coming later than I said, though, and AMI’s chief executive says it could take years for world models to move from research into commercial use.

Robots left the stage this year, and China led. About 19,100 humanoids shipped worldwide in the first half, nearly four times the year before, with Chinese makers accounting for over 97% of them. Unitree went from the Spring Festival Gala to a Shanghai listing in August where its shares rose nearly sixfold on the first day, with Reuters describing robotics as a key battleground in the US–China tech war. This is as much a machinepower story as a robotics one.

Autonomy is on the road too and arriving in the home. Waymo passed 500,000 paid rides a week in the US this spring, and 1X opened a factory in California to build its NEO home humanoid for its first US households, at $20,000 or $499 a month. Home robots are still early, though, and 1X says remote operators will step in for tasks NEO can’t yet manage alone, which is why I’m marking this one partly right.

Agentic commerce arrived through always-on personal agents, which run on their own cloud computers and buy for their users. Meta’s Muse passed 5 million downloads within three weeks of its September launch, and OpenAI’s Dots can shop too. These agents browse and check out on ordinary storefronts, so the year’s fights are about identification and permission.

Shopify has made its merchants purchasable inside Muse by default, with checkout settling over Google’s Universal Commerce Protocol. Spending is still small, at about $8bn of agentic transactions this year on Juniper’s estimate.

The live debates

• Is the agent you? In August an appeals court found that when Perplexity’s Comet shops on Amazon for someone, that person is the one accessing Amazon. Amazon blocked Muse anyway, saying it did not identify itself as an AI agent.
• Who gets to say no? Walmart partnered with Muse, Shopify opted its merchants in and Amazon shut it out. Perplexity says advertising is the real stake, and told the court Amazon was protecting the ads its human shoppers see.
• Who pays when it buys the wrong thing? Visa, Mastercard and American Express all run agent payment frameworks, but nobody has settled who carries the loss beyond plain fraud. Amex has gone furthest, committing to cover mistaken purchases by registered agents.
• Know Your Agent. The Treasury’s July AI plan calls for a trust framework with a Know Your Agent standard. Its payments consultation, which closed on 6 October, asks whether consent and liability rules need rewriting for agents.

Use and regulation both arrived. MIT Technology Review named AI companions one of its breakthrough technologies of 2026, citing research that 72% of US teenagers have used AI for companionship, and Oregon and Washington passed companion laws this year as nearly 100 chatbot bills were introduced across 34 states. The heart of the call, that emotional trust becomes a channel of influence, is still unproven, so I’m marking it partly right.

Copyright has moved into a licensing phase. Disney licensed more than 200 characters to Sora, and Universal and Warner settled with Udio on licensed services.

The clearest sign comes from the top. The President and the White House now use AI-generated images and video as a routine part of their communications, including a September video of Trump as Green Lantern, and Wired has called him America’s first generative AI president.

Generated software went mainstream this year. On 7 October OpenAI rolled out Intelligent UI to ChatGPT’s 1.2 billion weekly users, so answers can arrive as charts, forms and small tools built on the spot. Vibe coding did the rest, though its apps still go through the App Store, where new submissions rose 84% in the first quarter as people who had never written code began shipping software.