Author: Guy Wuollet, a16z Crypto

Translation: Deep Tide TechFlow

Deep Tide Introduction: Uber connected passengers and drivers with extremely simple expressions of demand, but many real-world transactions have been hard to digitize for a long time due to the involvement of multiple parties, dynamic budgets, and complex terms. AI agents are changing this situation, allowing systems to reorganize bilateral matching networks around more fragmented, non-standard goods and services.

The a16z Crypto team combines the mature "intents" concept in the cryptocurrency space with order flow bidding models to dissect the profitability mechanisms and cold start paths of agent business networks: In an era where software development approaches zero cost, machine networks that can accumulate fulfillment credibility and bidirectional traffic are the true long-term barriers.

Uber provided an extremely simple way for people to express demand: taking a ride from location A to location B. The network then automatically matches a driver willing to provide pick-up and drop-off services at a price you agree upon. The more drivers there are, the easier it is for passengers to get a ride; the more passengers there are, the more incentive drivers have to join the network.

I believe AI agents will allow such bilateral networks to be established around an extremely wide range of goods and services. Agents can help humans express more complex preferences and find suppliers willing to precisely meet those preferences.

Some of these trading scenarios actually already exist but have kept most ordinary people out due to high prices or extremely cumbersome collaborations; others have simply not been attempted because the costs of manual arrangements are too high. Agents will also spawn many new types of transactions that we have never even thought of before, creating huge opportunities for entrepreneurs.

Imagine a simple and familiar daily scenario: organizing a dinner with a group of friends. Each person's availability, budget limit, and restaurant preferences differ.

Today, a friend often has to reluctantly act as the "coordinator" in a WeChat group: collecting everyone's requests one by one, searching for restaurants across the internet, and renegotiating repeatedly in the group when the selected restaurant does not meet the conditions. Meanwhile, many restaurants may have plenty of open tables on Tuesday nights but lack effective channels to reach groups planning a dinner.

If handled by your personal agent, it could quickly align with your friends' agents to summarize a clear request: find a restaurant in the Mission District for a dinner of six, at an average of under $60, any time between 5:30 PM and 8 PM, as long as sushi is not included.

Subsequently, each restaurant's agent can immediately price based on available tables: some may offer better discounts for the early slot at 5:30 PM, while another could proactively customize a family-style meal suitable for six people. Your agent would just need to compare all quotes and everyone’s preferences and automatically complete the reservation within the pre-authorized range.

This way, restaurants with available tables at 6 PM can directly put out offers to consumers willing to dine earlier before they have selected where to eat. The buying team has clear preferences and budgets, and the selling restaurants possess available tables and price flexibility; the only thing missing is the underlying network that enables both agents to discover each other and complete transactions.

Uber has proven how networks can facilitate transactions around relatively simple travel demands, and agents are expected to expand this model to dining scenarios that have complex constraints. Most internet networks were built around human operating habits in the past, but now we urgently need to build networks for machines.

Entrepreneurs can definitely start creating dedicated matchmaking hubs that allow personal agents to bring complex demands to the market, letting suppliers compete for fulfilling these demands.

How to Build an Agent Trading Network

In fact, the cryptocurrency field has already provided a mature organizational paradigm around the concept of "intents": Users only need to declare the final desired trading outcome, and various service providers will bid against each other to deliver that outcome.

Taking UniswapX as an example, users sign an order that clearly states the tokens they want to exchange and the acceptable terms; market-making solvers then compete across multiple sources of liquidity to facilitate the transaction.

AI agents can extend this "intent model" to everyday commercial fields. Large language models (LLMs) can convert everyday natural language dialogues into standardized requests that suppliers can execute, while agents take responsibility for searching all network quotes and finalizing transaction details. This low-barrier interaction is expected to reduce the previously cumbersome collaboration costs, making small-to-medium transactions like a dinner for six financially viable.

The same network effects that once supported Uber's value will also gradually accumulate in these more complex transaction scenarios. Each time a new restaurant joins, diners gain a precise matching option, particularly valuable for small groups with strict dietary restrictions or time constraints.

As the number of groups initiating demands increases, restaurants gain more opportunities for additional table turnover income, which in turn will force them to integrate their reservation systems into the network and issue more targeted quotes. Once the participation level in a single-city market reaches a critical point, this network can naturally expand into more consumer categories.

In terms of cold start strategies, entrepreneurs can start with a single city's dining scene. The initial product must truly save users significant time or money to motivate them to bring in friends; simultaneously, partnered restaurants must see pure incremental orders they normally wouldn't receive. Only when both sides taste the benefits will they continue to stay and invite new members.join.

As for commercial monetization, I believe that agents aimed at end consumers can be completely free, while making money through a model similar to "Payment for Order Flow." Specifically, users declare what they want and how much they are willing to pay at most, then allow various suppliers to compete to meet their needs.

Suppose the agent quotes $300 for a dinner and successfully matches a restaurant willing to accept the order for $270, the network platform can keep the price difference or charge a routing commission for successfully directing the transaction to the restaurant. This matchmaking revenue can then directly subsidize the user's AI inference cost.

Of course, the platform must ensure transparency: users should clearly understand how their agents make money and the underlying logic for sorting through multiple quotes. Ultimately, the willingness of consumers to delegate spending decisions depends on their level of trust in this filtering mechanism.

Why This Type of Business Model has Long-term Barriers

As artificial intelligence significantly reduces the marginal cost of software duplication, simply having useful software is no longer rare; the network of individuals and business entities willing to settle within a specific network and continuously engage in transactions is becoming a more core competitive moat. Reflecting on the internet era, the greatest companies like Facebook, eBay, and Uber all followed this rule: the more participants that join, the more useful the entire network becomes for individuals.

A matchmaking network for agent business can similarly create this unshakeable first-mover advantage. Competitors may quickly replicate software with extremely similar functions and interfaces, but they still must gather reliable buyers and sellers who can fulfill transactions. Returning to the restaurant example, this means covering a sufficient number of local restaurants willing to make proactive offers and a large group of consumers ready to place orders at any time.

Guaranteeing that every matched transaction can be reliably executed is a key part of building network infrastructure. Partnered restaurants must be assured that the agent making requests has actually secured payment authorization from the entire dining group; similarly, diners need to be confident that once they click to accept a quote, they can reliably lock in a table at the agreed price.

Therefore, the network must build a set of verification mechanisms to authenticate agent permissions, record core terms of reservations, and smoothly arrange fund settlements. Additionally, in cases where a group cancels a reservation last minute or a restaurant fails to hold a reservation as agreed, both sides should clearly understand what the pre-agreed mechanism for handling such exceptions is.

Entrepreneurs can definitely build these trust facilities alongside the trading market. As soon as buyers and sellers begin to interact frequently within the network, the platform can naturally accumulate a set of credible records, marking who can reliably fulfill commitments and establishing mature dispute arbitration processes. A network that can help participants discover quality quotes and fulfill transactions reliably will naturally encourage everyone to return for future dealings.

What Other Agent Networks Can Founders Build

Restaurant reservations are just a simplified example I use to illustrate the idea. In reality, AI agents can assist in building numerous distinctive vertical trading networks, and the following directions are particularly worth exploring:

· Travel and Business Trips: Personal agents can align the schedules and expense budgets of multiple travel companions, then negotiate bundled pricing with multiple hotels and airlines, finalizing an itinerary that satisfies everyone. When teams maintain flexibility around departure dates or destination choices, they can often unlock hidden discounts that traditional fixed keyword searches cannot reach.

· Enterprise Procurement: An exclusive procurement agent for a company can be online 24/7, dynamically negotiating long-term supply quantities and price lock-in periods with upstream suppliers. When comparing prices, it can simultaneously weigh delivery schedules and payment terms, helping procurement parties secure total costs while providing sellers with predictable order demand.

· Computing Power Resource Market: Agents can automatically match computing task loads with idle computing power across the network based on hardware specifications, geographical regions of data centers, and time windows. Companies can even use this network to purchase forward computing capacity in advance to hedge against soaring computing costs during business expansion periods.

· Online Dispute Arbitration: Transaction parties can agree upon signing that any disputes will be submitted to a shared arbitration forum, which will resolve them far more efficiently than the traditional judicial system. Each time a new set of buyers and sellers adopt the same arbitration rules, the forum's attractiveness and credibility for subsequent market participants will rise correspondingly.

· Physical Device Networks: Agents can directly purchase transportation services from distributed drone delivery networks or acquire real-time data from connected sensor devices. Every new device that connects directly expands the coverage radius of the physical network or overlays new service capabilities that other ecosystem participants may urgently need.

· Collaborative Health Research: Individual users can securely contribute health data to a shared pool and participate in crowdsourced research projects. The larger the pool of data contributors, the more likely it is to support unprecedentedly large clinical control studies, thereby tackling medical challenges that previous small sample groups could not answer.

Additionally, around "real human relationships," there are also rich opportunities for network entrepreneurship. As machine intelligence becomes increasingly cheap, even close to free, real society still urgently needs effective means to verify real identities, endorse professional qualifications, and build deep trust relationships. As more work moves away from traditional employment agencies, networks that provide professional reputation credentials, membership clubs, and similar organizations will become even more valuable.

Conclusion

Real business opportunities often lie in the transactions people genuinely wish to reach; however, the time and money barriers for manually coordinating these transactions are exceptionally high today. The largest types of agent networks in the future are almost certain to arise from entirely new exchange scenarios that we currently cannot imagine.

By entering a specific and nuanced vertical market, first running through bilateral demand fulfillment, and then incentivizing each participant on both ends with tangible incremental value, we can bring more friends into the fold. If you are working on building agent networks, I look forward to discussing this with you.