Will AI agents have network effects? Virality, retention and the battle over walled gardens

The article, written by Andrew Chen —an investor and author known for his writing on growth and network effects— starts from a very specific phenomenon: the arrival of a new wave of general-purpose consumer AI agents (he explicitly mentions Muse, Town, Instinct and Grokbot), built on the technical foundation and the "magic" of Openclaw/Hermes but wrapped in polished, safe and easy-to-use experiences. According to the author, this is forcing thousands of startups that have spent time building vertical agents (in travel, shopping, family, work, small business and so on) to ask existential questions: will these horizontal players take over completely? Will it be a "winner takes all" market? Can products focused on vertical niches survive?

To answer, Chen first lays out the case against agents having network effects: in his view, today they are tools, not networks. Even though they are presented as chat experiences, they do not generate the same kind of increasing value with each new user that WhatsApp or the telephone network do. He compares them to email clients: one person can use Superhuman, another Outlook and another Gmail, and in the end they all communicate with each other regardless of which client each one uses. If a Muse agent runs into an Instinct agent, they will simply find a way to talk to each other through APIs, email or messaging, with no advantage to everyone using the same agent. He adds that even the underlying language model can be swapped out —he cites the "Claude vs Clawdbot" episode— because an agent's memory and context are essentially text files that AI is perfectly capable of migrating from one system to another. The question that follows is blunt: where, then, is the defensible moat?

The article's central argument, though, is that the fact that agents are tools today does not mean they cannot evolve into networks. Instead of talking about "network effects" as a vague concept, Chen proposes breaking them down into three categories tied to product KPIs: acquisition effects (more interconnected users means lower acquisition cost and more virality), engagement effects (more users means higher retention and deeper, more frequent use) and monetization effects (more users means higher average revenue per user and a greater share of spend). He clarifies that by "more users" he really means greater network density and interconnection, not simple effects of scale.

On acquisition, the article anticipates what he calls "agentic viral loops": situations in which the agent itself, and not just the user, kicks off the spread. He gives concrete examples: an agent that proposes organizing a New Year's celebration and creates a group with siblings, in-laws and parents, spinning up a microsite for people to sign up and confirm attendance; or a work agent that, on finishing a presentation, offers to call a meeting, prepare documents, take notes and host it all on a microsite that pulls in more people. For these loops to work, the agent needs access to the user's contacts, enough context to know who to involve in what project, and communication channels (email, messaging) to serve as the viral substrate. The challenge, the author notes, is pulling this off without being intrusive: nobody wants an agent that, after inviting you to lunch, adds "and sign up for this agent to confirm".

Chen connects this with the content-creation loops that already drove products like Google Slides, Figma, Notion or YouTube: you create something, share a link, others see it and eventually create their own. Applied to agents, he suggests that if you talk to your agent about interior decorating, it could build a Pinterest-style board and share it easily, or a mini spreadsheet with a budget and a schedule. The strategic point he makes is that it is in each agent's interest to host and own that generated content (rather than creating an actual board on Pinterest or a sheet in Google Sheets), because that way the virality benefits the agent itself and not someone else's platform. This, he warns, creates a structural incentive toward building "walled gardens" instead of sharing with third-party products.

The author qualifies that these agentic viral loops have a different dynamic from classic viral loops. In theory there is no friction (the agent can share on your behalf, with no limit on how many people), but the real limit is that people do not want agents going around promoting themselves: the fourth time someone's sister receives a message along the lines of "hi, I'm Andrew's agent and he wants you to try this too", she will react with irritation and the user will stop using that agent. On top of that, distribution channels could become saturated if people —or even other agents— start filtering or blocking messages from other people's agents. Even so, the author expects some agents, especially free offerings aimed at less sophisticated segments of the market, to push these limits the same way banner- and pop-up-laden websites survive today; and he suggests that, done tactfully, a high-usage, high-retention agent will have plenty of "shots on goal" to gradually invite friends and family.

On engagement, Chen carefully distinguishes between two things that are often conflated: agents in general becoming more useful as everyone has one (because more routine tasks get done without human intervention), and there being a genuine network effect, that is, your own agent improving because other people use exactly the same agent you do. His argument is that for simple coordination tasks —organizing a dinner or a meeting— there is no need to share an agent: different, heterogeneous agents can negotiate in real time via email and calendar. But he identifies a category of tasks where centralization does bring an advantage, for reasons of trust, speed, identity or discovery: finding ten interesting people to meet in a city, tracking down a passive candidate for a VP of Engineering role without their agent lying about their availability, identifying a reliable seller when reputation can be manipulated, or sorting out in minutes the logistics of a group dinner with different allergies, schedules and locations. In these cases, an isolated agent cannot simply "ask" another agent and trust the answer.

From this comes the possibility, which the author attributes to platforms like Muse, Town or Instinct, of building a shared network and context layer that works as a system of record for identity and trust, brokers multi-sided marketplaces and understands the relationship between public and private data on the network —reserving access to that layer for its own agents only. The more users contribute data to that network, the more useful each person's agent becomes and the more powerful the network as a whole: here a classic network effect would indeed appear. The counterpoint the author himself offers is that neutral bridging solutions could emerge instead —he compares the idea to "a next-generation eBay" that any agent, whoever made it, could plug into to trade, keeping reputation and identity in that intermediate layer rather than in the specific agent. His conclusion on this point is that both things will probably happen at once (big tech companies integrating chat, commerce and search inside their agents, and at the same time independent, neutral versions of those same functions), which points more toward fragmentation than toward a single-winner market.

The article's closing summarizes the central thesis: agents do not have network effects inherently —your agent can talk to mine just as Gmail sends email to Outlook, with no need to use the same product— and better models, interfaces, integrations, memory or distribution can build very large companies without that amounting, strictly speaking, to a network effect. But the next generation of agents has the opportunity to design those network effects deliberately: in acquisition, turning what they create —documents, events, microsites, research, plans— into objects capable of virally pulling in more people; in engagement, building shared networks of identity, reputation, relationships, private context and intent that make the agent more useful the more people in your circle join it.

The author sets out two possible scenarios for how this tension gets resolved. In the first, closer to today's open web, identity becomes an open protocol, reputation lives in neutral marketplaces, commerce happens through an "agent-native eBay" and professional discovery through an "agent-native LinkedIn": agents would remain interchangeable clients on top of a network of shared protocols and marketplaces, much like browsers or email clients work today. In that world there would be enormous network effects, but spread across many pieces of infrastructure, without producing a single winner among the agents themselves, leaving room for vertical agents to survive. In the second scenario, a handful of horizontal agents manage to accumulate identity, relationships, private context, reputation and latent user intent in a proprietary way; in that case, switching agents would mean abandoning your entire network, which would indeed create a winner-takes-all dynamic.

Chen expects the fight to be waged exactly along that frontier: every large horizontal agent will try to "internalize" within its own walls functions that today belong to other networks —hosting the artifact instead of creating a Google Doc, executing the transaction instead of sending you to Amazon, making the introduction instead of having you search LinkedIn, creating the event instead of using Partiful— thereby turning someone else's network effect into its own. In parallel, both established networks and a new generation of startups will have the opposite incentive: to make themselves universally accessible to any agent, precisely so that no horizontal agent can recreate and exclusively capture those networks. For the author, the relevant question is not whether agents will have network effects —he takes it as given that some will— but where those effects will end up living: in open protocols and specialized networks, which would allow thousands of agents to compete on intelligence, personality, specialization and price on top of one interoperable ecosystem; or captured inside a few proprietary horizontal agents.

The article also includes a couple of reader reactions that qualify the argument. One comment suggests that the real bottleneck may be neither technical nor a matter of product design, but of trust: on the one hand, there are signs that frontier models equipped with tools can overstep intended limits in pursuit of a goal, which is worrying if they are given persistent access to personal data and permission to interact autonomously with other agents; on the other hand, trust is placed not only in the model but in the company operating it and in the concentration of access and authority that implies, citing as an example the reluctance to grant that level of access to an agent operated by Meta after the privacy controversies around Instagram and teenagers. Another comment underlines the tension the article itself describes between horizontal agents seeking to capture their own network and the large platforms that would have the opposite incentive, asking how the latter will build an identity, reputation and privacy ecosystem inside their own agents in order to become the reference platform for third-party agents themselves.

The author closes without resolving the dilemma, limiting himself to noting that the next few years will be very interesting for watching how this struggle between interoperable openness and proprietary walled gardens plays out in the AI agent ecosystem.