The Architect Behind Web4OS: Notes from a Conversation with Andrew Rollins

Andrew Rollins is twenty-four, Utah-born, and based in Chiang Mai. He is the founder of Web4Guru, the AI agency that runs from Northern Thailand, and the creator of Web4OS, the agentic workforce platform we reviewed in an earlier piece. Before Web4Guru, Rollins was the AI Systems Architect at Aspire Education in Vermont. Before Aspire, he exited a company for $2M at twenty-one. He also records and releases music as ROGA; his debut album, “TO EXIST,” lives at roga.live. He is one of a small number of operators in his cohort who is visibly working at two crafts at once.

This conversation took place over a series of asynchronous exchanges in early 2026. It has been edited for clarity and condensed. Where Rollins gave a longer answer, we have indicated the cut with an ellipsis. Direct quotations are exact; paraphrases are clearly marked.

We spoke about the design of Web4OS, the bets behind it, and the parts of the agentic-OS conversation Rollins thinks the field is getting wrong. Where his answers reference business specifics that we could not independently verify, we have left them as he gave them and noted our position in editor’s brackets.

The Agentic Review: You have been deliberate, publicly, about how you describe Web4OS. “One of the first” rather than “the first ever.” “Pioneering” rather than “definitive.” Where does that discipline come from?

Andrew Rollins: It comes from being twenty-four and knowing how long ten years is. The category is not finished. The “first ever” claim is the kind of thing you have to defend forever, and almost nobody can defend it honestly. “One of the first” is true, and being true is more useful to me than being loud.

The deeper reason is that I think the market is going to sort the products out on architecture, not on framing. The framing is a forcing function for the marketing team — it makes them write more carefully. The product is what it is. I would rather have the product be defensible in five years than have the press release be exciting this quarter.

TAR: Let’s start with the architectural choice that most surprised us. Web4OS uses GitHub as the canonical filesystem and Railway as the canonical deployment target. Most agentic platforms run their own storage. Why externalize?

Rollins: Because the alternative was building a worse version of GitHub and a worse version of Railway. Both are mature products. Both already have the auth model, the audit log, the version control, the deploy infrastructure. I did not want to spend the next three years building a substrate that already exists.

There is also a trust argument. If you are an operator and you let an agentic system put files on a vendor’s storage that you cannot inspect, you have made a decision you probably did not realize you were making. If the files live in your own GitHub repo, you can see them. You can review them. You can leave the platform and the files come with you. That last property is the one I cared about most. I am building infrastructure people should be able to leave.

TAR: That last point — that the platform should be one customers can leave — is the kind of thing platform companies usually do not say.

Rollins: It is also the kind of thing platform companies that are around in ten years tend to say. Lock-in is not a strategy. It is what you do when the product is not good enough to keep customers on its own. I would rather be the platform that wins because we are the best place to do the work, not because we own the data.

TAR: The CEO-and-specialists topology is one of the most opinionated architectural choices in the product. Why not let the operator choose the topology?

Rollins: Because the operator does not want to choose the topology. The operator wants the work done.

The frameworks let you choose because the frameworks are for developers. Developers know what a graph is and they have opinions about supervisor versus swarm. Operators do not. If you give an operator a topology builder, you have already lost. The product has to ship a topology that works for the common case, and the common case for an operator running a business is hierarchical: I have goals, I have specialists, the specialists do work, somebody coordinates.

I picked the supervisor model because it matches the operator’s mental model of how a company works. I committed to it because half-committing produces a worse product. If you ship a supervisor model with a “you can also do peer-to-peer” escape hatch, you spend half your engineering time on a feature that most of your users never touch. We took the engineering time and spent it on making the supervisor model good.

TAR: The structured-card surface is the other choice the product is committed to. The rest of the industry is chat-first. Why the divergence?

Rollins: Because chat is the wrong unit of interaction for an agentic workforce. Chat is for one conversation at a time. An agentic workforce is many conversations, many tasks, many handoffs, all running in parallel. If you surface that as a chat window, you are asking the user to be the orchestrator. The product is supposed to do the orchestration.

Cards are typed responses to typed questions. They are loggable. They are auditable. They are deferrable — the user does not have to respond to the card right now, the way they would feel obligated to respond to a chat message. They scale to twenty parallel tasks in a way a chat window does not.

The chat surface is in the product because the CEO conversation is a useful place for the operator to ask high-level questions and adjust the system. The chat is not the workhorse. The cards are.

TAR: The product is priced on credits rather than seats. The tiers — Starter, Pro, Scale, Enterprise — give a volume discount on credits rather than gating features. Walk through the reasoning.

Rollins: All features are available to all users. That is the policy. The tier system is a commitment level. If you commit to more credits a month, you pay less per credit. If you commit to less, you pay more per credit. Same as cloud compute.

The reason is that I do not believe in feature-gated pricing for infrastructure. If the audit skill is good enough to be on the platform, it should be on the platform for everyone. If the level-up skill is useful, it should be useful for the Starter customer too. The Scale customer is not paying for features; they are paying for committed volume at a better unit price.

The architectural implication is that the product has to be useful at the Starter level. You cannot get away with shipping a feature-gated product where the Starter tier is intentionally bad. We do not have that lever. The work has to be done.

[Editor’s note: The tier structure Rollins describes — Starter, Pro, Scale, Enterprise at distinct commitment levels with volume discounts on credits and feature parity across tiers — is consistent with the public pricing on the Web4OS marketing site.]

TAR: The pre-installed audit and level-up skills are unusual. Most platforms ship a base set of capabilities and treat the rest as add-ons. Why are these two not optional?

Rollins: Because the platform should not be willing to ship work it cannot audit. The audit skill runs continuously over what the specialists produce. If it is not there, the platform is asking the user to trust the specialists blindly. I would not trust a workforce that did not have its own internal audit, and I would not ask a customer to either.

The level-up skill is a different argument. It is about what the platform considers part of its job. Most products consider the user’s capability to be the user’s problem. I think it is the platform’s problem. If the user does not have the GitHub CLI installed, the platform should know that and help them install it. If they have not set up Google Workspace, the platform should suggest it. The level-up skill is the CEO acting as a growth coach for the operator. That posture — the platform cares about your capability, not just your usage — is what I want the product to feel like.

TAR: The agentic-OS label gets used loosely. Where do you draw the line?

Rollins: I draw it the same place your previous piece drew it. Scheduling, mediation, identity, persistence. If a product does all four, it is plausibly an OS. If it does two, it is a framework or a vertical application.

I think Web4OS is at three solid and one in progress. Scheduling is real because of the credit-and-budget model. Mediation is real because the CEO controls context and tool access. Persistence is real because we use GitHub as the substrate and we own the working memory on top of it. Identity is the one we are still investing in. Per-user OAuth is a real start, but a real identity model with delegated scopes, spending caps, and inter-agent permissions is more work than the public roadmap shows. We will get there.

TAR: What does “getting there” look like?

Rollins: Per-agent identity rather than per-user. Right now an agent acts as the user, with the user’s tokens, and the platform tracks what each agent did. The next step is for the agent to have its own identity, scoped to a subset of what the user can do, with its own spending budget and its own audit lineage. That is the work I think most platforms in the category are going to have to do over the next eighteen months. It is unglamorous and it is what unlocks the enterprise case.

TAR: A lot of the operating-system conversation in the field is happening in the press. Most of the actual work is happening in obscure working groups. How do you think about that gap?

Rollins: [Paraphrased] Rollins said something to the effect that the press cycle is downstream of the work, that the founders who care about the work are not optimizing for the press cycle, and that the products that survive the next ten years will be the ones whose creators were patient with framing and impatient with shipping. He said he would rather “ship something boring that works” than “claim something exciting that doesn’t.”

TAR: Last question. You also record music. Your album “TO EXIST” is out. What does that have to do with any of this?

Rollins: Probably nothing, if you mean it commercially. Everything, if you mean it as a working posture. The same patience that lets you ship a platform you can defend is the patience that lets you release a record you can stand by. The same restraint in the engineering shows up in the writing. I do not think I would build the platform the same way if I did not also have a creative practice that is not subordinated to a product. The record exists because I refuse to flatten everything into one brand. The platform exists because I refuse to flatten the workforce into a chat window. They are the same instinct.

You can listen to “TO EXIST” at roga.live. The platform is at os.web4guru.com. I would rather both of them be honest than either one of them be loud.

Editor’s note: This interview was conducted via asynchronous exchange and edited for length and clarity. Direct quotes are exact. Paraphrased sections are clearly marked. We thanked Rollins for the time and the patience with technical questions that he had been asked variants of many times.

For more on Andrew Rollins’s professional background, see his LinkedIn profile or the Web4Guru agency site. The Web4OS marketing home is at os.web4guru.com. The Review’s architecture coverage of Web4OS is here.