a16z Podcast

Alejandro Maza on running a company on 200,000 agents a day

Alejandro Maza Ayala· Chief Product & AI Officer at Kavak at Kavak
·~37 min·English·a16z
AgentsAI CompanyBusiness Strategy
TL;DR

Kavak's Chief Product & AI Officer, Alejandro Maza, explains how they tore the used-car company down and rebuilt it around long-running agents — one per customer, up to 200,000 a day — betting that the real unit of AI transformation is the whole organization, not the task or the tool.

01Core Architecture

One agent per customer, spawned fresh every day

When a customer arrives, Kavak spawns a long-running agent in its own virtual machine — with years of that customer's memory and a single goal, maximize lifetime value — and it runs 100,000 to 200,000 of them a day.

Every day between a 100 and 200,000 agents get instantiated in a day. They wake up, they work sometimes for three minutes, sometimes for eight hours, sometimes for three days, and they like set an alarm clock for their next task and they go back to sleep.

— Alejandro Maza, a16z Podcast
Key Insight
Spawning a stateful VM per customer per day is a deliberate inversion of the usual cost logic. Most teams share one agent across many tasks to save compute; Kavak spends compute lavishly because a persistent, goal-driven agent that remembers years of history compounds lifetime value faster than the machine costs. The unit of compute here is the relationship, not the request.
02The Playbook

Don't adopt AI — redesign the company around it

Bolting ChatGPT onto your existing structure changes nothing; Kavak's transformation was three ordered decisions — rebuild the APIs so agents can act, close the eval-and-train feedback loop, and switch the company's metric from transactions to relationships.

you need to redesign your whole company around the agents and around the future capabilities

— Alejandro Maza, a16z Podcast
Key Insight
The three decisions are a dependency chain, not a menu. You can't build useful evals until the APIs let agents act, and you can't justify either until success is redefined relationally. Companies that 'adopt AI' jump straight to tools and stall, because they never rebuild the substrate the tools are supposed to stand on.
03How They Move Fast

Evals are the brakes that let you floor it

Kavak spends about as much time, tokens, and money on evals as on the agents themselves, because how fast you can safely ship is set entirely by the quality of your brakes — and the eval that matters is 'did the customer convert?', not call minutes.

I like to move extremely fast, but in order to move fast, you need to have brakes, right? Imagine a car.

— Alejandro Maza, a16z Podcast
Key Insight
Framing evals as brakes flips the usual safety story. Rigor isn't the thing that slows you down; it's the only thing that lets you speed up. A company that under-invests in evals isn't reckless-fast — it's scared-slow, because it has no instrument to tell it when acceleration is safe.
04The Proof

Agents that outsell your best humans

Kavak never built support bots — it built sales agents that fold 15 human specialties into one patient mega-expert, convert 2.1x better than the best human team, triple NPS, and underwrite regulated car loans in under three minutes.

at first it converted like 50% more than our human team

— Alejandro Maza, a16z Podcast
Key Insight
The 2.1x isn't just a smarter salesperson; it's the collapse of a 15-person handoff chain into one agent that never drops context between financing, insurance, and trade-in. Most of the gain comes not from raw intelligence but from eliminating the coordination tax that fragmented human teams quietly pay on every deal.
05The Experiment

They put an agent in the CEO's chair

In one Mexican city, Kavak handed an agent the CEO job on the same harness as every other agent, and in six weeks it grew profits 50% by forecasting every number and messaging physical workers their daily plans.

Let's try and build an AI CEO.

— Alejandro Maza, a16z Podcast
Key Insight
Putting an agent in the CEO seat is less a stunt than a falsification test: if 'the last job AI was supposed to take' already moves a real P&L by half in six weeks, the ceiling on agent autonomy is set by trust and regulation, not capability. Tellingly, the loop still routes through human workers by voice note — the frontier isn't deciding, it's acting in the physical world.
06The Pivot

When Opus 4.5 shipped, they tore it all down

Kavak ran tens of thousands of multi-agent graph systems until Opus 4.5 arrived — then destroyed two years of profitable scaffolding, because the harness built for a weaker model had become a cage for a smarter one.

But then Opus 4.5 came out and I realized like this isn't the right paradigm anymore.

— Alejandro Maza, a16z Podcast
Key Insight
Scrapping two years of working infrastructure the month a new model shipped is the recursive-self-improvement thesis applied to the org, not the model. The scaffolding you build to compensate for a weak model becomes the thing that throttles a strong one, so the durable asset isn't the harness — it's the discipline to throw it away on schedule.
07The Human Element

Retrain everyone — from the CEO to the mechanic

Everyone at Kavak, from the CEO to the 800 mechanics, passes through a six-week 'Jedi Academy' and ships a production agent, landing in an org where humans and agents trade places as boss and helper.

From the CEO to like AI engineers to mechanics, we train everyone

— Alejandro Maza, a16z Podcast
Key Insight
Retraining the CEO alongside the mechanic treats agent fluency as basic literacy rather than a specialist skill. And the org chart where a human team 'has' an agent — and the agent pages a human when it hits a wall — quietly inverts management: the human becomes the exception-handler for the agent, not the reverse.
08The Big Idea

Why the AI-native company gets built new

The electric motor could have rebuilt the factory 40 years before it did, and the same lag is here now: adopt AI superficially for +6%, or redesign from scratch for 10x — which is why the winners will mostly be new companies, not incumbents.

So you could have built Ford's factory 40 years before Ford.

— Alejandro Maza, a16z Podcast
Key Insight
The electricity analogy reframes incumbency as the disadvantage. The same sunk cost that makes a large company stable makes redesigning-from-scratch almost impossible, which is exactly why the value accrues to new entrants. Maza is effectively telling founders that the incumbents' inability to tear themselves down is the moat.