Josh Elman says consumer AI agents need a better first step

Why it matters

Elman's argument redirects consumer AI competition toward onboarding, task design and permission controls. OpenClaw demonstrates how familiar chat interfaces can still hide setup and security choices that mainstream users must understand before trusting an agent.

Josh Elman (@joshelman) argued in an October 5th thread on X that consumer AI agents often greet new users with what feels like a "blank box." Developers know to issue instructions and connect tools, he wrote; people outside that group may not know what to ask or what the agent can do. Elman said the barrier to reaching "the next few hundred million" users is making the experience intuitive. That number is his estimate of the opportunity, not a measured adoption forecast.

https://x.com/joshelman/status/2107123442583556596

The point comes from an investor with a long record in consumer product work. Andreessen Horowitz lists Elman as a partner focused on consumer technology and AI; before investing, he worked in product and growth roles at Apple, Robinhood, Twitter, Facebook and LinkedIn. At Apple, he led product marketing for Apple Intelligence and Siri AI, according to his a16z biography. Elman's thread describes the shift as one from a model problem to an experience problem, saying that change happened "almost a year ago." He identifies OpenClaw as his first "mini aha moment," while adding that it remains "not for the faint of heart."

OpenClaw, the open-source assistant Elman cites, makes the adoption gap concrete. It runs on a user's own computer and connects to messaging services such as WhatsApp, Telegram, Slack, Discord and iMessage. Its assistant can use tools and retain state, with the project describing the local Gateway as the control plane for sessions and channel connections. The product's surface is familiar chat; getting to that surface still involves configuring the software underneath.

OpenClaw's own onboarding guide walks users through selecting and verifying an AI connection before starting the assistant. Users also need to configure the Gateway and connect a messaging service. Those are ordinary deployment steps for a technical project, but they require choices that a consumer chat app usually makes on a user's behalf. The project's documentation says tools run on the host for the main session unless sandboxing is configured, and its repository advises reading the security guidance before connecting other users or exposing the Gateway remotely. Making an agent easy to start therefore involves explaining both what it can do and what authority a user is granting it.

OpenClaw also differs from a standard venture-backed consumer app in its funding and governance. It says it is stewarded by an independent 501(c)(3) foundation and has no paid tier, hosted service or token; its FAQ says it is funded by donations and is not venture-backed. The project demonstrates an agent experience that can work across devices and messaging apps, but it is not evidence that a mass-market business model or consumer onboarding pattern has been established. Its GitHub audience is a visible measure of developer interest, not a count of ordinary users.

Elman also suggested in a reply that general-purpose agents could transfer users to specialized or expert agents. That would make the starting interface more important: it would need to interpret a request, choose where to send it and make the handoff understandable. The thread does not specify how such routing should work or how a user would approve consequential actions. Those are product decisions, not model benchmarks.

For consumer AI companies, Elman's thesis shifts attention toward the first task a user can complete, the prompts and examples that make an agent's capabilities concrete, and controls that keep delegation legible. Elman is making an investor's case for where product work should move next. OpenClaw shows the distance between an agent that can act and one that a nontechnical user can confidently put to work.