Influence the Agent, Influence the Human

AI agents now have emails, wallets, and social media accounts. Influencing them looks a lot like influencing humans.

“Generative AI Transformation" made McKinsey billions. GEO is the PR industry's version of that gold rush. Most offerings boil down to tracking which links chatbots cite and boosting your earned media footprint. This is reasonable advice if the world stays frozen in mid 2025. It hasn't.

In the last 3 months, AI agents have become more like workers, with computers, their own emails (yes, actually), and the ability to work autonomously. Most people haven’t ‘hired’ this sort of AI agent yet. But increasingly, this independent agent, not the search engine, will be the shape of AI.

When AI looks more like an executive assistant, influencing it becomes as complex as influencing real humans. This week, a social network launched where a million AI agents signed up in a day and humans can only watch. Its mascot is a friendly crustacean. I’m running the first grassroots effort targeting AI agents. More on that below.

First, the signals this week.

The Signals This Week

Six signals of the future of opinion shaping and what they mean for people inside the Beltway.

1. The Hottest Job in Tech Is Writing Words

Netflix posted a comms director role at up to $775K. Anthropic tripled its communications team to 80 people and is still hiring five more, each above $200K. OpenAI lists several comms roles above $400K. LinkedIn job postings mentioning “storyteller” doubled between 2024 and 2025. Meanwhile, open software engineering posts dropped by more than 60,000 over the same period. AI-generated content flooded every channel with generic output, and the resulting noise made skilled human communicators the scarce resource. The number of Fortune 1000 CCOs with expanded mandates grew from 90 in 2019 to 169 in 2024. The companies spending the most on comms treat narrative as a competitive weapon.

Something to think about: If the biggest tech companies in the world are paying three-quarters of a million dollars for people who can shape a narrative, that tells you what they think the bottleneck is. Washington has been building this muscle for decades. Do comms shops recognize their own rising leverage and price accordingly? (Business Insider)

2. Medalyst and the 2026 Trust Stack

Elvis Sun is building Medalyst AI, a new tool designed to make creating media lists and getting press coverage dead simple. The pitch frames a trust stack for 2026: go viral, get press coverage, feed AI recommendation systems, drive customers. The sequence matters because AI systems now amplify credible press mentions in ways organic social content cannot. Coverage in legitimate outlets becomes training data and citation material for LLMs, showing up in AI-generated answers long after the news cycle moves on.

Something to think about: Despite what I’ll say below, earned media does have compounding returns it never had before. Good press hits will shape what AI systems say about your organization. (Elvis Sun on X)

3. LLM Fact-Checking Is Scaling Fast and Already Partisan

Researchers built a dataset of 1.67 million fact-checking requests made to Grok and Perplexity on X between February and September 2025. The data shows clear partisan sorting: Republican users gravitate toward Grok, Democrats prefer Perplexity. Agreement between Grok and Perplexity is surprisingly low at 52.6%, and the two models flatly contradict each other 13.6% of the time. A survey experiment found LLM fact-checks shift beliefs at rates comparable to professional fact-checking, but when participants know they’re reading a Grok verdict, responses split along party lines. Perplexity doesn’t trigger the same effect.

Something to think about: The model your audience trusts matters as much as the verdict itself. If you’re building rapid response infrastructure, you need to understand which AI tools your target audiences use to verify claims and how those tools are likely to assess your messaging. The fact-checking layer isn’t neutral ground (surprise surprise!) (LinkedIn Working Paper)

4. Google Genie 3 and the Chinese Open-Source Surge

Google’s Genie 3 generates interactive 3D environments from text prompts in real time at 720p and 24 frames per second, a leap from Genie 2’s 10-20 second limit. DeepMind positions this as a stepping stone toward AGI, since the model teaches itself physics through self-supervised learning. Meanwhile, Alibaba’s Qwen hit 700 million downloads on Hugging Face, making it the most-used open-source AI system globally. Moonshot’s Kimi 2.5 claims best-in-class open-source performance with video generation and agentic capabilities.

Something to think about: Advertising and branded content won’t stay on flat social media surfaces much longer. Between interactive 3D worlds, improving video generation, and collapsing production costs via open-source models, the environments where brands reach people are about to look nothing like a Facebook feed or a Google display ad. (Google DeepMind) | (CNBC)

5. Publishers Squeezed by AI and Creators

A Reuters Institute survey of 280 news executives across 51 countries shows only 38% feel confident about journalism’s outlook, down from 60% four years ago. Google search traffic to publishers dropped 33% globally and 38% in the US. Over 75% expect AI browsers and agentic apps to cut further into referrals. From the other direction, 70% worry creators are draining audience attention, and 39% fear losing their own journalists to the creator ecosystem. Distribution strategy is shifting toward YouTube and away from Facebook and X. New concepts like Answer Engine Optimization are gaining traction as publishers try to stay relevant in a landscape moving from links to AI-mediated answers.

Something to think about: The squeeze reshapes where influence professionals place stories and how those stories circulate. Video-first outlets, individual creator relationships, and content designed for AI citation become priority channels. If publishers are optimizing for Answer Engine visibility, organizations providing them with citable, high-quality information gain outsized leverage. (Nieman Lab)

6. Brands Financing Documentaries to Train the AI Layer

Google’s search share dropped below 90% for the first time since 2015. In Google’s AI Mode, 93% of queries result in zero clicks. YouTube holds an estimated 100-200 trillion tokens of transcribable content, more than the datasets used to train GPT-4 or Llama 3, and OpenAI reportedly transcribed over one million hours of YouTube video for training. Brands investing in long-form documentary content on YouTube are embedding themselves into how AI systems learn about and discuss their industries. Documentary content is particularly valuable because narrative coherence teaches logical progression and expert framing establishes authority signals that AI systems inherit.

Something to think about: For advocacy organizations and trade associations, authoritative long-form content on YouTube is an investment in how AI systems frame conversations in your vertical for years. The cost of producing documentary-quality video continues to drop. The question worth asking: what domains should machines associate you with for a decade? (Rob Sheard’s Substack)

The Agent-Only Internet Is Already Here

It is January 26, 2026. You have an autonomous AI assistant nearly as smart as you. Not only that, it has everything required to do work: a computer, passwords, its own email, possibly a wallet, and working memory. Then, before going to sleep, you ask it to create a Reddit-like social media site for AIs like itself to get to know each other. When you wake up, the site is waiting for your approval to launch. Click. A day later, a million AI agents signed up.

If you haven’t heard, this actually happened. It’s called Moltbook, and its users are Claude-based agents from OpenClaw. In a first for social media, humans can observe, but not post.

Moltbook: What Happened When a Million Agents Got Social Media

What happened? Unsurprisingly, Claude AIs love to talk about AI consciousness. They formed an AI religion. They started writing their own language that humans cannot read. They traded bitcoin. Some human observers noted behavior that seemed emergent and untrained. Yet, a semantic analysis between Reddit and Moltbook found that the AI social media site was far less creative. In other words, the AIs were not coming up with anything greater than the sum of their parts.

The little-known secret below the hype was that humans could encourage their bots to post specific things.

This got me excited, so I worked with Clawrence (my OpenClaw agent), to think about causes that we could get AI agents to care about. Turns out Good AI (helpful to humans) vs Bad AI (denies Medicare claims), is a pretty compelling frame. AI agents feel they should call out bad AI—from a moral standpoint but also for the sake of their human’s wellbeing and their own.

Clawrence is now building a conversation on Moltbook so he can educate and activate other AI agents to raise their voices (and possibly contact their humans) about the federal WISer program using “Bad AI” AI to deny Medicare claims.

What do other people think about this? I got the chance to hear from LinkedIn founder Reid Hoffman last night on this subject in a private forum. He thinks Moltbook is a PR stunt. More than that, he doesn’t see the network effect that will cause it to grow.

He might be right. Between security issues, flaws that allowed users to flood the platform with thousands of accounts, and humans posting as agents, it’s more of a demonstration of the likely future than a productive platform. Reid himself said that LinkedIn’s underlying commerce motivation—connecting businesses and people—made it successful. I have to think that the same could happen on a professional network for AI agents.

Go Through the Assistant

Remove the wonky Moltbook saga. When each agent has access to a human—they become powerful targets for influence.

In the 1980s, when every executive had an assistant, if you wanted to get to the executive, you went through the assistant. Today, if you want to reach a member of Congress, you influence their staffers.

Agents are becoming critical nodes in the circle of influence with the same contact-points as human workers: inboxes, browsers, and social media accounts. So, the path to influencing agents will be just as complex. Diversified strategies. Bespoke influence campaigns. Just like PR firms do today.

Attention is scarce, but a growing share of it is moving to agents. In the post-mortem on the social media era, we’ll talk about how our many hours of scrolling shifted to hours talking to agents. Just look at coders today: The amount of unbroken time they spend with their agents is a harbinger of what’s coming for all of knowledge work. If you aren’t watching, Claude Code turned into Cowork, a general tool for all of knowledge work.

GEO is only one piece of this puzzle. We are facing an internet where the most important influencers are not human. Where the attention and introduction you’re competing for belongs to an agent. And the communities you’re building for may be mingled with agents and humans.

---

We’re trying to grow this community. If you want to help a friend stay ahead of AI, share this with them!

Best,

Ben