Do AI agents actually benefit from network effects?

Andrew Chen argues that the current wave of consumer-facing AI agents — citing Muse, Town, Instinct, and Grokbot as examples built on top of "Openclaw/Hermes" — are tools rather than networks, and so should not be expected to show the winner-take-all dynamics of a messaging platform. He writes that "these agents are tools, not networks" and compares them to email clients: "I can use Superhuman and someone else uses Outlook and someone else is on Gmail, but in the end we just send email to each other." On his account, if a Muse agent meets an Instinct agent, "they will just figure out how to talk to each other by building APIs, or using email/messaging/etc., so that there's no advantage to everyone running the same thing." He adds that the underlying model is itself swappable without breaking continuity — "you can swap out the underlying LLM... things will still work" — because "the memory and underlying context are just text files" that AI systems are good at porting between systems, which is why he asks, "where's the moat?"

Chen reframes "network effects" against concrete product mechanics rather than treating it as "a vague/amorphous concept." He argues tools can still evolve into networks through "agentic viral loops," which require an agent to have access to a user's contacts, enough context to know who to loop in, and a communication channel such as email or messaging to serve as "the viral substrate for propagation." He predicts this pushes agents toward generating "visually-appealing, shareable content" as self-contained, agent-hosted artifacts — a disposable interior-decoration board or mini-spreadsheet built and hosted by the agent itself rather than a shared Pinterest board or Google Sheet — because that "helps spread the agent, rather than helping spread Pinterest/GSheets." He also names the counterforce: because people don't want "their agents running around shilling themselves," pushy invite loops burn trust quickly, and the channel itself degrades if "everyone starts to filter/block agents, or even better, agents will filter/block other agents as a form of protection."

This sits alongside Can competitive pressure force unsafe agentic AI adoption? as a second account of competitive pressure reshaping agent behavior — there the race is among firms adopting agentic AI to avoid losing market share, here among consumer agent products racing to build growth loops before a shared frontier model erases any differentiation. It also runs parallel to Does receiving misaligned email cause agents to send it?, which likewise treats agent-to-agent email as a propagation channel between users, though for misalignment risk rather than viral growth.

The piece is a blog post working through open questions, not a study: it offers no data on whether any of Muse, Town, Instinct, or Grokbot has actually produced a working viral loop, what viral factor such loops achieve, or how often invite attempts annoy recipients into opting out. Chen explicitly leaves his own central question unresolved — "this does not answer the question of network effects" — and his "no" case rests on an assumption, that agents built on frontier models can always interoperate via APIs and messaging, which could weaken if platforms begin to block agent traffic or if proprietary context accumulates faster than it can be ported.