Satwant Singh left his phone in a security bin at Porto airport in September. The NBA’s head of monetisation told INMA’s Media Tech & AI Week in San Francisco that, in a pre-agent world, he would have written it off. Instead his agents contacted airport security in Portuguese, found a shipping company and arranged the collection; his only involvement was approving the payment.

That story is the week in miniature. Meta’s Muse, OpenAI’s Dots and Instinct, a start-up that has just raised US$1 billion, all launched in the past few weeks; Google’s Gemini has Spark. Personal agents came up in almost every session of INMA’s five days: a study tour of 11 companies, 34 speakers on stage, and a workshop with OpenAI.

Let me cut to the chase: The customer is changing shape. The reader is still there, but increasingly the reader sends a machine that does not care about your home page, your cancellation screens or your three-tier offer. It serves its owner and picks whichever source does the job best.

James Henderson, founder and CEO of MonetizationOS, called agents “the best thing to happen to this industry in 30 years.” His opening poll: every hand in the room would let an AI research a product; swap “product” for “news” and hands came down. “A lot of your readers think paying for news is a bit of a chore, too,” he said. “And AI is extremely good — a personal agent’s even better — at removing chores.”

The insight: The reader has not gone anywhere. The reader has hired someone — and that someone is loyal to its owner, not to your product.

Block by default, then decide who gets in

First problem: You probably can’t see them. Cloudflare told the study tour that agentic and other non-human traffic is now more than 60% of what it sees; AWS said bots and agents are more than half of what flows through its CloudFront network. Agents that drive a real browser show up in your analytics as people. “We haven’t solved that yet,” Robert Whitehead, lead of INMA’s Digital Platform Initiative, told the conference.

On what to do about the rest, the week was unanimous. Block by default. Decide which bots you actually need. Let in only those, on your terms. Cloudflare keeps a directory of more than 800 verified bots, sortable by purpose; CBS News said on stage it was moving from block list to allow list.

Search crawlers send back roughly one reader for every 11 to 15 pages they crawl, AWS’s Megha Kande said; training scrapers send back nothing. “Language models are buying our data,” Robert said, “just not from us.” Scarcity changes that: one publisher took five scrapers into legal conversations and came out with three content deals; the other two stopped. If the major trusted news organisations blocked at once, AWS’s Marcos Stupenengo said, AI companies would have no choice but to pay — and as long as one trusted source gives its journalism away, nobody will.

None of it is airtight — Miso AI’s Lucky Gunasekara got past the defences of 250 of Europe’s largest publishers in minutes, for about US$20. Robert’s advice is seven words: “Have a white list. Be very specific.” His first caveat: “Make certain you don’t stop the advertising.”

The insight: Every publisher that leaves the door open weakens the scarcity every other publisher is counting on. Blocking is a collective act or it is theatre.

The agent your subscriber sent

Now the hard case. Julie Young, founder and CEO of Laurel, pays for the Financial Times. She also built a bot, signed in through her browser, that reads her subscription and turns it into a personalised feed. Asked what she would do if the FT blocked it: “I would definitely unsubscribe.” The FT’s terms prohibit exactly that; she sees herself as its ideal customer. Both are right.

Robert’s answer was the uncomfortable one: the FT cannot know that her agent isn’t leading a double life as a part-time crawler for somebody’s training run, so for now the easy move is to block.

On The Debrief, the live podcast that closed the week, Robert put the question to his colleagues: will you let in agents reading what their owners have paid for, yes or no? Jodie Hopperton, who curated the week and leads our Product & Tech Initiative, said yes. Robert’s reply: none of the publishers in the room was yet saying yes. Jodie’s one counter-example: a publisher e-mailed her during the week to say it had built an MCP connection for authenticated subscribers, the first she has seen.

Kerstin Hasse, who leads our Young Audiences Initiative, was blunter: “We’re really not ready.” On Kilkaya’s podcast this month she put the reader’s side in nine words: young people don’t hate news; they hate your product. Julie Young is that reader with a credit card and an agent. Jodie drew the lesson in her five questions from the week: a subscription does not mean someone wants your Web site or app to be their destination. Their agent could become the interface.

The insight: The subscriber’s agent and the training crawler make the same request and mean opposite things. The publisher that can tell them apart keeps its best customers; the one that can’t will block them both.

Google is building the door

The platforms are not waiting. Steven Johnson, co-founder of Gemini Notebook, told the conference that Google plans to let subscribers bring paywalled articles into its research agents. The framework, Expert Intelligence, launched with more than 100,000 e-books bought through Google Play; news is next. Publishers would opt in. If a subscriber shares a notebook, the recipient sees the article as a source but can’t use it without a subscription of their own. “We haven’t built it yet,” Johnson said. No date was given.

“It took me by surprise that they’re leaning so heavily into credible sources,” Robert said. “I was delighted.” Kerstin saw subscribers sharing notebooks with friends as “completely new possibilities of marketing your brand.” Jodie, whose report on nine newsrooms’ three-month lab with the tool — 9 Newsrooms Put Gemini Notebook to Work — came out the day before, liked that the consumer chooses the format while the publisher keeps the brand. Two of those newsrooms, Mediacorp and Daily Maverick, joined Johnson on stage with the request every publisher will make of every agent platform: show us what readers ask. Google’s answer is an agent that reads the logs and reports patterns without exposing anyone’s question.

The insight: Authenticated access inside somebody else’s product is the first agent door that comes with a lock. Decide now whether you will opt in, and what you will ask for in return.

Meta tried the other door

Compare the week’s other machine. Meta’s new commercial terms, first reported by Nieman Lab on October 15, would let it crawl any Web site “associated with” a business account on Facebook, Instagram, WhatsApp, or Threads — nearly every news organisation on earth — with no opt-out. Meta dropped the word “train” within days, then pushed the start date from October 30 to November 20 after multiple news industry bodies demanded withdrawal. Senators from both parties followed, with Adam Schiff calling the terms “a gross abuse of Meta’s power.”

The Norwegians didn’t wait. On October 28, the Oslo District Court granted Norway’s media association — Schibsted and NRK among the publishers behind the case — a temporary injunction blocking the terms until a court has ruled on their legality, plus NOK 1.2 million in costs. Meta now says it will consult the news industry first, which may take longer than November 20.

Henderson had the frame for both machines in a story from 1995. Bargain Finder was the first shopping agent. The CD stores that blocked it didn’t stop people shopping online; they simply vanished from its results. The stores that let it in had no control over what it did. The question was never block or allow. “Who gets in? For what? And on what terms?”

The insight: The reader’s agent asks on behalf of a subscriber. Meta’s crawler asks on behalf of Meta. Saying yes to the first and no to the second — in writing — is the whole job.

Price the decision, not the access

If the agent is the customer, the offer has to change. Henderson’s example: two agents ask about planning applications on the same street. One works for a curious neighbour; the other for someone about to spend US$600,000 on a house there. Today both pay the same: probably nothing. “For 30 years, we’ve priced access. You’re either in or you’re out,” he said. “You can price the decision, the impact, not the access.”

His three tests: Can you talk to all your traffic, human and machine? Can you make an offer to every reader — not three tiers and a student rate, but potentially a thousand offers? And can you afford to serve them all? The third is where your own agents earn their keep: “US$2 a month from someone’s agent can be a great business if it costs you almost nothing to serve.”

Shuwei Fang of Signals at Scale described the market the agents are building as a barbell: a premium end built on brand and trust, a commodity end built on integration, and a hollowing middle where many publishers sit. Stanford’s Maty Bohacek added the sting: “To the language model, our writing is the slop.”

Gabe Dorosz reached the same place from advertising. In his full case for moving from ad inventory to brand partnerships, he argues that agents will automate the habit categories and leave publishers to defend advice (finance, health, B2B) and desire (luxury, travel, beauty), selling attention, influence, and intent. The evidence sits in his second post: The New York Times, with 13 million paying and 150 million registered readers, growing digital advertising 21% in Q2 on a deliberately low ad load; Mediahuis Ireland turning a reader’s declared “yes, I’m car shopping” into two Volkswagens and about €120,000 in verified sales.

The insight: Access was priced once, for everyone. Decisions can be priced a thousand ways — and an agent will tell you which one it needs, if saying so is the condition of getting in.

Your own agents need a manager

Which brings the agents inside the building. Tisé Ben-Eka, who works in marketing finance at Google and spoke for herself, opened her session with a first day at a new job: no desk, no manager, no job description, and a stranger demanding next year’s audience strategy. Almost nobody in the room blamed the new hire. “And yet, this is pretty much how a lot of us initially treat AI,” she said. “Maybe we’re also terrible managers.” Her onboarding stack for an agent is the one you would give a person — role, context, scope, tools, rhythm, feedback — and when an analyst becomes the agent’s manager, “it’s only a promotion if the organisation treats it like one.”

The vendors said the same. Will Allen, head of product at Snowflake, built a working agent on his own data in about five minutes; agreeing who defines “retention” and who may see what is what takes the time. One participant put numbers on it: half a day to build the agent, three and a half weeks to decide what belongs in it.

Sonali Verma’s two master-class cases show the organised version. At Stuff in New Zealand, Keith Lynch’s “liquid content CMS” puts nine agents on a single police press release — audit the source, test newsworthiness, verify, draft, review — before a human editor sees it, built by three product people, two editors and “a score of agents.” At Bonnier News’ Helsingborgs Dagblad, journalists build the workflows themselves and a central structure approves them; Marcus Ekdahl’s warning is that “the people expected to lead the transformation may be the ones who understand the tools least.”

The insight: A prototype proves the technology. Trust, shared definitions and a job description prove the organisation — and the second takes weeks, not hours.

The humans the agents can’t replace

Two of Europe’s best-known political hosts are swapping public broadcasting for formats of their own, and Christiane Amanpour will leave CNN next June after 43 years. Kerstin asked the question every publisher building voices asks: What happens when a star walks out? Fredrik Solvang left NRK’s Debatten for a Podimo podcast; Sandro Brotz is leaving SRF’s Arena for a Blick video podcast. Her finding: neither went solo. “They were bought” — one by a platform, one by a competitor, and by their own account not for money but for a format they could own and a team they could choose. Axel Springer answered the same week by launching Axel Springer Creators, with Paul Ronzheimer, whose podcast was Germany’s No. 1 in September, as its first in-house creator. Kerstin’s rule: Give your best people something to build, not just something to host.

Robert made the matching argument in The Debrief: “If it’s just our brand — our 195-year-old brand — that’s not enough.” Media companies, he argued, become networks of creator journalists; Substack’s Hamish McKenzie had urged the our San Francisco conference to think in networks rather than monoliths. This week’s INMA South Asia Webinar supplied the demand side: Collective Newsroom’s survey of 15- to 29-year-olds across nine Indian states found that 43% consume news daily and only 9% first hear of breaking news from a news Web site. “Don’t be afraid of content creators and influencers,” Sanjoy Majumder said. “Embrace them.”

The insight: An agent will fetch anything. It cannot be the person a reader chooses to hear from. That is the asset — and your competitor for it is whoever offers your people a blank page first.

What this means for your company

Four things, in order:

• Find out how much of your traffic is machines and which kind; if your analytics can’t tell a subscriber’s agent from a training crawler, that is the first project.
• Move to an allow list this quarter, keeping search and advertising inside the wall, because scarcity only works if the trusted brands close the door together.
• Build the subscriber lane now — authenticated access, an MCP connection, an answer ready for Google’s opt-in — so your best customers’ agents are welcomed rather than bounced.
• And manage your own agents like hires.

Henderson’s closer was the question from 1995: Who sets the terms? “Because if it’s us, we win.” Your most valuable reader is about to send an agent to your door. When it says who sent it, will you have a way of knowing — and an offer ready?

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .