AI is closing the gap between 

• desire and action
• belief and knowledge 

In the near future, we’ll just be able to say what we want and our agent will get it done. 

But today we’re on a path of protectionism, where it’s implied it’s an American duty to pay high prices to a concentrated set of American companies for intelligence. 

Agents are too important to live behind a $20 / month subscription. 

The open internet broke free of private networks through innovations in security and data protocols that unlocked e-commerce. 

This provides a playbook to free and open agents. 

‍

E-commerce and memory unlocked the open internet 

The early internet was controlled by two private networks, AOL and MSN. They charged both users and businesses for access to the network. 

Engineers at Netscape believed in a different future: One of open networks controlled by no one. 

But for this belief to succeed, it needed an economic model – rather, a data and security model – to support it. Netscape’s Lou Montulli explained

The thought was that if we enabled e-commerce we would essentially create a revenue stream for the internet that would be sustainable [and outcompete private networks]

And one of the core technologies that we needed to create was some form of memory for the web so that when a user came to a site and expressed some sort of preference that when they continued through the site they would remember [their preference].

Cookies unlocked memory like clicks, scrolls and infrequent purchases, but within the pocket universe of a given site or app. 

But we expect agents to operate on our complete context – making the internet’s current memory model insufficient. 

Actions our agent may wish to take 

• Reading the NYT
• Searching Airbnb 
• Booking an Uber

are often guarded by apps – “no one in tech wants to be someone else’s dumb API call” Ben Evans wrote in January. 

But apps that own valuable actions don’t know our agent and its complete context. Their first party data is nothing compared to what our agent knows about us.   Montulli continues

Cookies were complex in that they were attempting to solve the data privacy problem while giving memory to websites. The core concept of cookies is what if we sandboxed them to only the relationship to a given website. 

The first step was giving the web memory through cookies and CDPs. 

With agents, the next step is giving the web context-aware delegated agency: Our agent needs a way to take action and apps need a secure way to call our agent. 

‍

Agents can already take reliable and cheap action

For our personal agent to be valuable, it needs to be able to take reliable and inexpensive action. 

While many have focused on browser-led agent action, many are finding the economics don’t work in the consumer case. 

Browser-use charges $0.05 per agent step. These economics can work in enterprise – but do they work in consumer? $0.25 extra to have your agent order DoorDash when the alternative is use the app for free? 

Every expense that agents incur takes us further away from free agents. Worse, folks find browser agents slow and unreliable. Sites are hiring engineers to block them. 

The browser path may not be needed, as agents can already take reliable action through APIs:

• Kroger
• Instacart
• eBay
• Hotel + flight APIs

already offer much more economical and purpose-built interfaces into action spaces consumers care about. While some apps may choose to eschew agents, others may (and already do!) see open interfaces into (agent-lead) commerce as valuable sales channels. 

Most of these API integrations are completely free, enable reliable and purpose built action, and need only someone to register a client id. 

This way, users can have their personal agent take action on their behalf reliably and for the lowest possible cost. Since the personal agent knows the user’s complete context, it doesn’t just call Kroger’s API naively, it does so with all available context. 

‍

Internet memory drives commerce but is expensive 

In 2023, McKinsey reported that 

companies that excel at personalization generate 40 percent more revenue from those activities than average players

Over three-quarters of consumers (76 percent) said that receiving personalized communications was a key factor in prompting their consideration of a brand, and 78 percent said such content made them more likely to repurchase.

But personalization is expensive. 

• Databricks ACV is around $150k
• Snowflake ACV is around $300k, 
• CDPs cost anywhere from $50k - $800k per year
• Machine learning engineers or data scientists cost around $200k / year 

But as we’ve written before, agents introduce new economics in personalization, making it better and cheaper. 

Can spending on personalization infrastructure more efficiently collapse into calls to a personal agent? 

‍

Agent API calls pay for free agents 

The free internet was paid for by ads. 

Can free agents be paid for by apps calling your agent? 

Apps usually want you to buy, engage, or return more often. 

Personal agents do “in each instance, exactly what the user would want it to do if asked specifically.”

Personal agents may not have all of the capabilities of apps, but they do know the user the best. As a user agent, they can also make sure that only the context that’s relevant to an app is actually shared with a purpose built interface. 

Personal agents enable apps to deliver experiences that are more relevant – encouraging higher spend, engagement, retention, and reduced cost of reacquisition. Personal agents can also save apps money, by reducing spend on expensive data and personalization infrastructure. 

And relative to shipping a remote Model Context Protocol server, personal agents have 

• unified context 
• streamlined consent screens 
• and trusted operations

that make agent context ops efficient for AI and intuitive and safe for users.  

‍

Agent alignment is about taste and economics

It’s generally accepted that personal agents should be aligned to users. 

At first pass, it seems like the only way to do that is to have users pay. 

We’re suspicious of this argument. We pay internet service providers and they’ve sold our data. A monopolist or member of an oligopoly may not be aligned to us either. 

If AI agents are going to be important, they should be available to everyone on the web regardless of ability to pay. 

Accepting that means defining new business models and pursuing capital structure that doesn’t induce misaligned incentives like ads, monopoly, or price increases when VC money runs dry. 

“Agents calling agents” has been a popular refrain in the Valley, but we believe their strongest motivation lives in economics and (cross-trust boundary) context management. 

There are of course interesting questions on what “your agent” is – we believe they’re more like an user agent headless server with a variety of AI client interfaces – than any singular interface or God model. 

This is the future we believe in – one where technology saves us time at the lowest possible cost. Crosshatch is the internet’s first personalized inference provider, providing the internet’s fastest way to log in with your personal AI enriched with complete context. 

If you’d like to experiment with this world – try talking to your agent on our playground or stay tuned for our Personal MCP Server to land and link it to your favorite MCP client.