Six takeaways and five stealable frameworks

Before diving deeper, here are the most valuable insights from this episode. Six takeaways:

Five frameworks you can copy verbatim

Details below. The ordering matters: the first few points are about building the foundation, the later ones about what happens when you apply that foundation to business and org structure.

Three foundational decisions: rewriting APIs, setting superhuman standards, changing KPIs

Everything stems from three decisions. None are about tech stack choices; they all answer the same question: What are we trying to become?

1. Redesign the company, don't just hand out tools

On an engineering level, this means rewriting most APIs and systems so agents can use them to get things done. It was a necessary step because interfaces people can tolerate are often unusable for agents. A workflow that requires a person to guess, click around, and copy-paste between three systems simply can't be executed by an agent. If they hadn't changed the interfaces, agents would've been stuck on the sidelines.

The next step was building the data and feedback loops needed to train these agents. How? By releasing them to face real customers, gathering data, getting eval results, and then training.

2. Bet on building superhuman agents.

"Superhuman" has a concrete definition here: on the dimensions that matter—conversion rate, customer lifetime value, customer experience—the agent must outperform the best human the company has ever hired. And it needs to be pointed at the hardest problems, not the easy scraps.

3. Shift from a transactional company to a relationship company.

Kavak used to be a transactional business, measured by cars bought, cars sold, and brake pads procured. Now it's a relationship company. There are 10 million customers in its database, most with an individual agent whose goal is to maximize that customer's lifetime value.

This shift works for Kavak because of two simple calculations. First, the money: they sell high-ticket items like cars and large personal loans. Activating just 1% of those 10 million people correctly is worth hundreds of millions of dollars. Second, the nature of the industry: buying a used car requires trust before a customer will spend. And trust is built by actually knowing the customer and managing the relationship long-term—which is precisely what these agents are good at.

One agent per customer, each with its own virtual machine

This is the most technically dense and unconventional part of the whole piece. While most teams are building "multi-agent systems" with various specialist agents, Kavak is betting on a different path: long-running agents with hard goals.

When a customer engages, the system spins up a dedicated agent for that customer, each with its own virtual machine. This agent remembers the customer's entire history with Kavak—what they viewed on the website, a phone call they made two years ago—everything is in its memory. It sets its own strategy, defines a long-term goal, and then figures out how to keep that customer happy and, at the right moment, sell them the company's various products.

Why they scrapped a profitable, two-year-old architecture

But this wasn't the architecture they started with.

By December 2025, Kavak had tens of thousands of agents running business on a "multi-agent system plus graph plus orchestration layer" architecture. A scheduler would break complex tasks into steps, dictating the order of operations and which model to call when. And it worked: it brought profitability and growth.

Then Opus 4.5 came out. Their assessment: the paradigm was no longer valid.

So they discarded the profitable, two-year-old system entirely and replaced it with the new model: one VM + one agent + memory + evals + CLI access to all company tools and APIs + a long-term goal. The benefit of this architecture is that it absorbs every new monthly model release: the stronger the model, the stronger the system. The old architecture had the opposite dynamic: the stronger the model, the more it got held back.

Evals Are as Important as Development

This aggressive approach only works because of evals. To drive fast, you need reliable brakes. You can floor the accelerator because you trust the brakes.

Many companies get this backwards. Because they lack good brakes, they drive slowly. Kavak flips it: how fast we can go depends on the quality of our evals.

Here's a rule you can copy directly: spend roughly the same amount of engineer time, tokens, and money on evals as you do on building the agents themselves. This is far more actionable than vague advice like "take evals seriously," because it's a budget ratio.

What to measure? Only these three things.

Business outcomes always come first. Is the customer happy? Did they buy a car, get a loan, or sell us their old one? The first check is simple: Did they convert?

Most companies fall apart here, measuring surface-level metrics like call volume or duration. Those have some information, but they don't solve the problem. You only need to ask three questions:

Once these three evals are in place, the rest is just optimizing the agent architecture, adding skills, and scaling to millions of customers.

AI Sales Outperform the Human Team

Kavak never built a customer support agent. The first agent they put in front of customers was the sales agent—the hardest and most valuable role in the company.

Why selling a car in Latin America is so hard

A customer has to pick from around 20,000 cars, go through financing, get insurance and extended warranty, and often sell their old car back, which means it needs to be appraised. This is what Kavak looked like in 2020 and 2021: you had to be good at 15 different things, so they had 15 teams of 15 specialists. The customer would start with a financing expert, then a car-selection advisor, then a trade-in specialist, then an insurance expert—they had to assemble a solution across all these people just to buy a single car.

Kavak's approach had two steps: first, they trained agents to beat the specialist in each domain, then they merged them into a "super specialist"—one agent that understands insurance, financing, car selection, and appraisal—and put it in front of the customer.

Why agents sell better

According to the interview, there are five reasons: they're experts, they have infinite patience, they know the customer's entire history, they can play the long game, and they never get tired. But the sixth reason is the most interesting:

There's another capability worth noting, stemming from their vertical integration: when a customer can't make their car loan payments, they can trade their car in for a cheaper one. This lowers their monthly payment and prevents financial ruin. Traditional auto lenders can't offer this—they don't have the cars, so they can only send collection agents.

The "AI CEO" Experiment

Going further up: Can AI handle management roles, even the CEO position? Their take: "Probably yes. At this rate of improvement, definitely by 2035." So they decided to test it now.

Their approach: designate a city in Mexico—Cuernavaca—and put an agent into the same harness to act as its CEO. It started learning, making decisions, and evaluating its own performance.

What it actually does

This agent acts like a Fields Medal-level genius: it goes through every number and every customer, makes predictions, and then micromanages every daily execution needed to make its plan a reality.

It's that concrete—it will directly message frontline workers in Cuernavaca about their daily plan and then ask them to send voice notes back with progress updates.

The result: several metrics improved simultaneously. Customer satisfaction went up, inventory quality improved, turnover sped up, and financial penetration increased.

People assumed the CEO was the last job AI would take. Turns out, that's not the case.

The Self-Improving Company | Kavak's AI Playbook · a16z

How Mechanics Use Their AI Sidekick

But some jobs are still done by humans. The roles they're still hiring and training for are those tied to the physical world, with mechanics being the prime example. This work requires a high degree of manual dexterity and sensory judgment, making it very difficult to replace.

So here, the agent is a sidekick, using the same harness. They named it El Mike (phonetic). It tells the technician how to inspect a car, offers tips, and demonstrates the process.

The results: significantly improved inspection quality, faster inspections, faster repairs, and lower costs. And most importantly, the cars being delivered are of higher quality, which gave customer satisfaction another boost.

So these roles are still being hired for and trained, but AI acts as an amplifier.

Company-Wide AI Training: From Executives to Mechanics

Where do they find people who fit this system? They train them themselves.

Kavak has an internal training program called Jedi Academy, for everyone from the CEO to AI engineers to mechanics. The curriculum is designed and taught by Maza himself and has to be constantly updated. You can't just send people to Stanford for this—it's entirely new.

The message to employees is simple: there's no turning back. The company is going in this direction, and they've been clear about what that means for the engineering, finance, and product teams. You can choose to learn, gain the skills to work in this new reality, or you can leave. The result is a stronger culture, with employees genuinely excited about the direction.

The org structure has changed, too

Kavak is now a very flat, very senior, very empowered team. Looking at a team, you'll see engineering, AI, and operations all mixed together. These teams are either building agents, working for agents, or facing customers in the physical world.

When Agents Hit a Wall, Humans Catch Them

This flips the reporting structure between humans and agents. It's easy to misread this part.

First, let's look at the common approach. "Human in the loop" in most large-scale agent systems usually means: when an agent hits a wall, they escalate the case or customer to level-2 human support, and then forget about it.

Here's the problem: The loop is never closed. A human might solve this particular ticket, but the system gets no data from it. The next time the agent hits the same wall, it hits it again.

Kavak reversed the direction: the agent focused on a specific customer has access to all the APIs and skills (built for it by humans and other agents). When it hits a wall or needs to cancel something, it calls an API saying "I need help," and the entity that catches it is a human—not another agent, not a piece of software.

On an org chart, it looks like this: a human team is assigned an agent, but the agent is the one giving orders. This arrangement solves a specific engineering problem: if the loop never closes, you never get the data you need to improve the agent.

A Three-Tier Framework for Evaluating Token ROI

So, how do you decide if this investment is worth it? He offered a framework for self-assessment, answering the question, "I'm spending hundreds of millions of dollars a year on tokens—what am I actually getting for it?"

Kavak is currently at Tier 3, which is why they can keep scaling up: knowing the ROI of each token turns expansion into a data-driven decision, not a gamble. When you know the return on each token, scaling is justified, not a leap of faith.

I spend more on tokens today than I do on knowledge workers.

The Self-Improving Company | Kavak's AI Playbook · a16z

The conclusion: it comes down to having a clear vision and a way to measure whether every token you spend is moving you toward it—then iterating relentlessly. Adoption rates alone mean nothing.

Two prerequisites for companies that want to follow suit

First, it must be top-down. Just pushing for "adoption" won't get you anywhere. Expecting employees to spontaneously decide what to build and create something genuinely useful for the company—that kind of taste and strategy doesn't emerge from the bottom up. Leadership has to use it themselves and have a crystal-clear plan: define what the company should look like in three to five years, then build toward it, directing the team vertically toward that goal.

Second, measure what actually matters. That's the eval system described earlier, plus this token tier system.

The Organization Itself Should Be the Self-Improving Entity

If you connect all these practices, you get one flipped narrative.

Everyone talks about recursive self-improvement (RSI) in terms of a model making itself smarter. This thesis changes the subject:

Why the opportunity is with new companies, not incumbents

Here, he brings up Joseph Schumpeter's creative destruction: innovation hits the economy when old companies stay the same, while new entrants armed with new technologies displace them. Expecting existing companies to fully adopt new tech is historically rare. In the short term, this destroys value; in the long term, it's better for everyone, as these more efficient new companies provide better products and services.

Why this time is no different: because deep AI adoption is really hard. A CEO of a large, especially public, company who stands up and says, "I'm betting everything on AI, I'm going to tear down everything built over the last 40 years and rebuild as an AI-native company"—how many people would do that? So while they "adopt" AI at the margins, new companies can grow from day one with AI strengths as their foundation.

It happened with electricity, too

Kavak often cites the electricity revolution. The technology needed for Ford's assembly line was developed in 1879 and 1881. Edison had already started commercial power generation and built highly efficient dynamos. In other words, Ford's factory could have been built 40 years earlier. The tech was all there.

But the initial adoption of electricity followed a different path: companies kept their existing four-story, line-shaft-and-belt-driven factories and merely replaced the coal engine with an electric motor. It was a solid improvement—about a 6% efficiency gain.

The real play was to tear down that factory and rebuild it on a flat greenfield site—not in downtown New York, but in Connecticut or New Jersey. Design the entire factory around small electric motors and electricity from the ground up. That's how you get a 3x productivity increase, and that's what underpinned 20th-century America.

The computer era saw the same pattern, and it's happening again today: people want to adopt it, but they don't want to redesign the entire company, so they just adopt it shallowly. The result is a 6% or 10% improvement, not a 10x one. In other words, the innovator's dilemma is playing out again, just at an industry-wide scale.

His final message to founders: For the first time ever, anyone can access the world's most powerful tools and intelligence for practically free, or for $20 a month. So go build something. But build it deeply. You don't even need to assume exponential growth—just assume AI keeps improving at a linear rate, and build for that future.

The mainstream approach is to assemble a system of specialist agents. Kavak does the opposite: for each customer, it creates a dedicated, long-running agent with its own VM (an independent computer serving just that customer). This agent remembers years of history, sets its own strategy, and defines its own long-term goal.

Behind this architecture are three decisions:

• Rewrite the APIs: Agents can't use messy human-centric interfaces. Without this, they're just bystanders.
• Set superhuman standards: Outperform the best humans on conversion, LTV, and CX, on the hardest problems.
• Change the KPIs: Move from "how many cars sold" to 10 million customers, each with an agent, maximizing lifetime value.

This architecture went through one major dismantling. In December 2025, Kavak had tens of thousands of agents running a profitable, growing business on a multi-agent system with an orchestration layer (a scheduler that breaks tasks into steps and plans for the model). When Opus 4.5 launched, they decided the paradigm was wrong: once the model is smart enough, the orchestration layer becomes a ceiling, not scaffolding.

Sales was the first job Kavak gave to agents, and it's also the hardest: customers pick from ~20,000 cars, navigate financing, insurance, and trade-in appraisals. In 2020–21, that meant 15 things, 15 specialist teams. Kavak's approach: train agents to beat each specialist, then combine them into a "super specialist" to face the customer.

Evals only track three things—don't copy most companies and track call duration:

• Did this customer convert?
• Did we deliver value?
• Will they come back?

To decide whether to keep spending, Kavak uses a three-tier token framework: Tier 1 is employees using chat tools—value unclear, where most companies sit. Tier 2 is engineers in the codebase—can be estimated indirectly. Tier 3 is agents executing core business—ROI per token is directly calculable. Kavak is at this tier.