Own your own agents

SaaS was the appetizer. Services is the six-trillion-dollar main course.

For twenty years the tech industry believed one thing. Take a service, turn it into a product, charge per seat, scale through distribution. That built a six hundred and fifty billion dollar software industry. It also trained everyone to confuse tools with outcomes.

Customers never wanted the tool. They wanted what the tool was supposed to deliver. Now with AI software can deliver the outcome directly. No interface. No training. Just the result.

The entire SaaS era was a detour. We spent twenty years selling access to outcomes instead of delivering them.

Agent Layer Is the New Control Point

Remember the scene in The Matrix where Trinity needs to fly a helicopter. She does not take lessons. She does not read a manual. She calls the operator. “I need a pilot program for a B-212 helicopter.” Seconds later she can fly. She never sees the software. She never knows who made it. She just gets the skill.

That is enterprise software now. The old enterprise provisioned applications by role. Sales got Salesforce. Marketing got HubSpot. Each tool required training, configuration, cognitive load. The new enterprise provisions capabilities by sentence. The FP&A team gets a skill that pulls from NetSuite and formats reports the way the CFO wants them. No training. No interface. No brand.

No brand.

You ask your agent to track expenses. The agent finds a skill that does it. You never see the tool. You never know it exists. The tool becomes invisible plumbing. Systems of record do not die. Something still has to be canonical underneath. But you do not use Salesforce. You use a skill that calls the Salesforce API. The brand layer collapses. Worse than losing distribution, the SaaS company loses the customer relationship.

OpenAI’s Frontier diagram puts “Your systems of record” at the very bottom of the stack. Rendered in white. Almost invisible. Above it sit Business Context, Agent Execution, Evaluation and Optimization, then Agents, then Interfaces. The systems of record layer is white. You can almost miss it.

A company valued at over half a trillion dollars telling you, in a diagram, that “system of record” is plumbing.

And they are not just drawing diagrams. Anthropic’s Forward Deployed Engineers embed with strategic customers. Goldman Sachs gave Anthropic six months inside their accounting and compliance workflows. That knowledge does not stay at Goldman. It feeds back into Anthropic’s systems. The labs see which domains drive heavy API usage, which apps gain traction, which features deliver results. They are reading the map that their own customers drew for them.

What happens when the labs raise prices post-IPO? When twenty cents of inference becomes two dollars because investors demand margins? A founder I know built a vertical agent at $1,000 a month. Six months later the base models improved and a $20 tool could do most of the job. The product survived. The pricing did not.

Building your core business on rented intelligence is like building a restaurant where someone else controls the price of every ingredient. Works beautifully until the supplier opens their own restaurant next door.

That inference rent is not pricing risk. It is landlord economics.

Whoever owns the agent owns the customer. Lose the agent layer and you are selling pipes.

This Is Capital, Not Operations

Everyone is in on AI, some have gone far, many have not.

General Catalyst has been buying legacy services businesses and injecting AI. Crescendo acquired call centers and took gross margins to sixty to sixty-five percent while tripling resolved calls in the first week. Long Lake acquired eighteen businesses and drove twenty-five to thirty percent productivity gains. Dwelly bought six property management agencies in the UK and doubled EBITDA margins where their technology is fully deployed. GC calls it the Rule of 60. Buy the relationship, inject agents, expand what you deliver.

Vista Equity Partners manages over a hundred billion across ninety-plus enterprise software companies. Built what they call an agentic factory. Removing humans from the workflow. Target is the Rule of 70. Replace the people with agents, keep the customer, margins explode.

GC buys the services business to get the customer. Vista owns the software and stretches it until it becomes the service. Same destination.

Both are thinking operationally. How do I make this business better with AI?

Different question. What am I actually buying when I deploy money into this space?

SaaS and software was about buying and renting products. Agents is capital.

SaaS charged per seat. Scaled through distribution. Defensibility was switching cost of moving data. Every dollar was a bet on owning the tool. Agentic Capital is different. Every dollar is a bet on owning the agent layer between the customer and the outcome. The relationship. You can export data from any tool. You cannot export a relationship.

GC’s Rule of 60 optimizes for efficiency. Vista’s Rule of 70 optimizes for margins. Neither asks who owns the agent. The firm that owns the agent layer owns the customer. The firm that rents it is one pricing change away from losing everything.

Efficiency is a feature. Ownership is the business.

Own the Harness, the Model, and the Memory

A doctor you have seen for twenty years. She knows your history, your family patterns, your anxieties, the medication you tried in 2014 that made you dizzy. You could switch to a new doctor who graduated top of her class. Might even be smarter. But she does not know you. That understanding is not in your medical records. It is in the relationship. Twenty years. Cannot be transferred.

Same thing happens when an agent sits in the execution path of your business long enough. It accumulates institutional memory. Not the data in your CRM. The other stuff. The exceptions. The overrides. The reason your sales team always gives healthcare companies an extra ten percent. The precedent from the deal you structured last quarter that everyone references but nobody documented. The escalation pattern your support lead uses that is not in any runbook.

That knowledge lives in Slack threads, deal desk conversations, people’s heads. When someone leaves the company it walks out the door with them. An agent that captures those decision traces while doing real work builds something no competitor can replicate by copying features.

You can copy a product in a weekend. You cannot copy twenty years of knowing someone.

Owning the agent means three things.

The harness. How you orchestrate, deploy, and connect the agent to your workflows. Get this wrong and the agent is a demo.

The model. Building your own model sounds radical. Less radical than it looks. You do not need to compete with Anthropic or OpenAI on general intelligence. A model trained on your workflows, your institutional memory, your decision traces. No foundation model will ever see that data. Rent the model and you are one pricing change away from losing your margins.

The memory. The accumulated understanding of how your organization actually works. Compounds over time. The longer the agent sits in the execution path, the more irreplaceable it becomes.

The labs want to be the engine, the car, the driver, the road, and the toll booth. They send Forward Deployed Engineers to learn your workflows. Then they build it themselves. If you do not own your own memory you will be eaten by those who do.

Own all three and you own the customer.

SaaS was the appetizer. Services is the main course. The agent is the waiter who remembers exactly how you like your food.

The question is simple. Own the waiter, or rent your fate.