If the future of the internet involves a bazaar of agents paying each other for services, crypto will find a level of mainstream product-market fit it could previously only dream of. While I feel confident agents will pay each other for services, it’s less clear to me whether the bazaar approach will win. 

By “bazaar,” I mean a decentralized, permissionless ecosystem of independently developed, loosely coordinated agents — an internet more like an open marketplace than a centrally planned system. The canonical example of a bazaar that “won” is Linux. This contrasts with the “cathedral” model: tightly controlled, vertically integrated services managed by a few large players. The canonical example here is Windows. (The term comes from Eric Raymond’s classic essay, “The Cathedral and the Bazaar,” which framed open-source development as chaotic but adaptive — an evolutionary system that can outperform carefully curated structures over time.) 

Let’s unpack each condition — agentic payments and the rise of the bazaar — and then explain why, if both come true, crypto becomes not just useful, but necessary.

Condition #1: Payments will be integrated into most agent transactions.

The internet as we know it subsidizes costs by selling ads based on how many human eyeballs see an app’s page. But in a world of agents, humans won’t be going to websites anymore for online services. And apps will increasingly be agent-based instead of UI-based. 

Agents do not have eyeballs to sell ads to, so there is a strong case that apps will need to shift their monetization strategy to charge agents directly for their services. This is basically the way things occur right now with APIs — services like LinkedIn are free, but if you want to use the API (the “bot” user), you must pay for it.

Given this, it seems likely that payments will be integrated into most agent transactions. Agents will offer services and charge users/agents in microtransactions. For example, you may ask your personal agent to find a great candidate for a job on LinkedIn. The personal agent will talk to the LinkedIn Recruiting Agent, which charges a fee upfront for the service.

Condition #2: Users will rely on agents with hyper-specialized prompting/data/tools built by independent developers, forming a bazaar of untrusted agents that call on each other for services.

This condition makes sense in principle, but I’m not sure how it will play out in practice.

Here’s the argument for why the bazaar will form: 

• Right now, humans perform the vast majority of service work, and we go to the internet to solve discrete tasks. But the scope of tasks we delegate to technology is going to expand dramatically with the rise of agents. Users will need agents with specialized prompting, tool calls, and data to perform their specific tasks. The set of tasks will be too diverse for a small group of trusted companies to feasibly cover, similar to how the iPhone relies on a vast ecosystem of third-party developers to reach its full potential. 

• Independent agent developers will fill this role, empowered to create specialized agents by a combination of extremely low development costs (e.g., vibe coding) and open-source models. This will create a long tail of agents offering hyper-specific data/prompting/tools, forming the bazaar. Users will ask agents to perform tasks, and those agents will call on other agents with the specialized capabilities to complete them — who, in turn, will call on others — forming long daisy chains.

In this bazaar scenario, the vast majority of agents offering their services will be relatively untrusted because they will be offered by obscure developers and will be niche in their usage. It will be very difficult for the agents in the long tail to build the sufficient reputation needed to earn the imprimatur of trust. This trust issue will be particularly acute under the daisy chain paradigm, where a user’s trust weakens along each link in the chain as services are delegated further and further from the agent the user trusts (or can even reasonably identify).

However, when thinking of how this might be realized in practice, there are a number of open questions:

• Let’s start with specialized data as a major use case for agents in the bazaar and look at an example to ground ourselves. Imagine a small law firm that does a lot of deal work for crypto clients. The firm has hundreds of copies of negotiated term sheets. If you are a crypto company doing your series seed financing, you can imagine how an agent using a model fine-tuned on these term sheets could be very useful for telling you whether your term sheet is market.  

• But thinking through it more deeply, is it really in the law firm’s interest to provide inference on this data via an agent? Offering this service to the masses as an API effectively commoditizes the law firm’s data when what it really wants is to upcharge you for its lawyers’ time. And what about legal/regulatory considerations? The juiciest data usually has legal regimes that require it to be kept under lock and key — that’s a big part of why it’s valuable and why ChatGPT does not have access to it. But the law firm is highly restricted from sharing this data under its duty of confidentiality. Even though the underlying data is not being shared directly, I’m highly skeptical that the “fog” of a neural network is enough to make the law firm comfortable that information is not being leaked. Given all this, wouldn’t it be in the law firm’s interest to just use that model internally to offer better legal services than its competitors and continue to sell the lawyers’ time?

• It seems to me the “sweet spot” for specialized data and agents is when there is highly valuable data produced by a non-sensitive business (so not healthcare, legal, etc.) that is auxiliary to the core service it charges for. As an example, a shipping company (a non-sensitive business) produces a lot of valuable data as a byproduct of its shipping business (I’m guessing; I don’t really know the first thing about shipping). So this shipping company would probably be fine offering an agent that taps into this data for a fee, because it’s exhaust that is otherwise going to waste. This data is probably really valuable to a set of people (like maybe a hedge fund). But how many such scenarios exist? (Not a rhetorical question; if you know good scenarios, please message me.)

• On prompting and tool calls, I’m just unsure of what independent developers will offer here that isn’t mainstream enough to just be productized by trusted brands. My simple mental model is that if it’s a prompting/tool call that is valuable enough for an independent developer to monetize, wouldn’t a trusted brand just step in to build a business off of that? I think this is just a lack of imagination on my part — the long tail of niche codebases on GitHub offers a good analogy of what this could look like with agents. I welcome thoughts on great examples for use cases.

If the practical realities do not support the bazaar scenario, the vast majority of agents offering their services will be relatively trusted because they will be developed by major brands. Agents can restrict their interactions to a curated set of trusted agents, relying on trust chains to enforce service guarantees.

If the internet becomes a bazaar of specialized but largely untrusted agents (Condition #2) performing services for payment (Condition #1), then crypto’s role becomes much clearer: it provides the guarantee needed to underwrite transactions in a low-trust environment.

While users will interact with online services with reckless abandon when it’s free (because the worst that happens is wasted time), when money is on the line, users demand assurance that they’ll get what they pay for. Today, users get that assurance through a “trust-but-verify” flow. You trust the counterparty or platform you are paying for a service and verify you received the service ex post.

But in a bazaar of agents, trust and ex post verification will not be nearly as available.

• Trust. As established above, it will be very difficult for the agents in the long tail to build sufficient reputation for other agents to trust them. 

• Ex Post Verification. Agents will call on other agents in long daisy chains, so the ability for a user to manually verify work and identify which agent dropped the ball or acted nefariously will be significantly more challenging

The upshot is that the “trust-but-verify” paradigm we currently rely on will not be sustainable in this universe. And this is the precise environment in which crypto excels — exchanging value in untrusted environments. Crypto does this by replacing trust, reputation, and after-the-fact human verification with cryptographic and cryptoeconomic guarantees. 

• Cryptographic: The agent performing the service only gets paid when it can cryptographically prove to the agent requesting the service that it did the thing it said it would do. For example, an agent can provide a TEE attestation or zkTLS proof (provided we can get it cheap/fast enough) that it scraped data from a certain website, ran a certain model, or contributed a certain amount of compute. This is all deterministic work that is relatively easy to cryptographically verify. 

• Cryptoeconomic: The agent performing the service will stake an asset and be slashed if it is caught cheating, which creates an economic incentive to act honestly even without trust. For example, an agent can research a topic and provide a report — but how do we know if it did a “good job”? This is a much harder form of verifiability because it’s not deterministic, and getting fuzzy verifiability right has long been a holy grail of crypto projects. But I’m hopeful we are at the point where fuzzy verifiability will finally be possible by using AI as a neutral arbiter. We can imagine a dispute resolution/slashing process run by a committee of AIs in a trust-minimized environment (like a TEE). When one agent disputes another agent’s work, each AI on the committee can be given the inputs to the agent’s work, its output, and details about the agent (history of past disputes/work on the network, etc.). They can then make a call about whether to slash it. This will function as a form of optimistic verifiability, where economic incentives will prevent parties from cheating in the first instance.

Practically, crypto allows us to make payments atomic with proof of service — no agent gets paid unless the work is verifiably done. In a permissionless agent economy, this is the only scalable way to ensure reliability at the edge.

To summarize, if the vast majority of agent transactions do not involve payment (meaning Condition #1 is not met) or are with trusted brands (meaning Condition #2 is not met) we probably won’t need crypto rails for agents. This is because users are fine interacting with untrusted parties when money is not on the line, and when money is on the line agents can simply whitelist a limited number of trusted brands/institutions to interact with, and chains of trust can enforce the promises of services each agent is offering.

But if both conditions are met, crypto becomes indispensable infrastructure as the only scalable way to verify work and enforce payments in a low-trust, permissionless environment. Crypto gives the bazaar the tools to outcompete the cathedral.

Thank you to Zach (Axiom), cwm (Soulgraph), Felix (EdenLayer), ilemi (Herd), Lincoln (Coinbase), Nima (EigenLayer), and Tommy (Delphi) for their thoughtful feedback and discussion on this article.

Thank you to my colleague Jack for countless hours of debate on this topic. 

All information contained herein is for general information purposes only. It does not constitute investment advice or a recommendation or solicitation to buy or sell any investment and should not be used in the evaluation of the merits of making any investment decision. It should not be relied upon for accounting, legal or tax advice or investment recommendations. You should consult your own advisers as to legal, business, tax, and other related matters concerning any investment. None of the opinions or positions provided herein are intended to be treated as legal advice or to create an attorney-client relationship. Certain information contained in here has been obtained from third-party sources, including from portfolio companies of funds managed by Variant. While taken from sources believed to be reliable, Variant has not independently verified such information. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by Variant, and there can be no assurance that the investments will be profitable or that other investments made in the future will have similar characteristics or results. A list of investments made by funds managed by Variant (excluding investments for which the issuer has not provided permission for Variant to disclose publicly as well as unannounced investments in publicly traded digital assets) is available at https://variant.fund/portfolio. Variant makes no representations about the enduring accuracy of the information or its appropriateness for a given situation. This post reflects the current opinions of the authors and is not made on behalf of Variant or its Clients and does not necessarily reflect the opinions of Variant, its General Partners, its affiliates, advisors or individuals associated with Variant. The opinions reflected herein are subject to change without being updated. All liability with respect to actions taken or not taken based on the contents of the information contained herein are hereby expressly disclaimed. The content of this post is provided "as is;" no representations are made that the content is error-free.