Remember when the entire self-driving conversation lived inside disengagement reports? Every few months a spreadsheet would drop, everyone would argue about miles-per-intervention, and the winner of the argument was whoever had the most patient statistician. That era treated autonomy as a perception-and-planning problem. Solve the driving, the thinking went, and deployment follows automatically.
It didn’t. Waymo is now running roughly 450,000 paid rides a week, double what it was doing not long ago, with plans to enter 12 new cities by 2026 including Dallas and San Diego. And the news that arguably matters most to that expansion isn’t a model update. It’s a lobbying budget.
Policy as a runtime dependency
Those of us who work on agent architecture tend to draw the system boundary at the software: sensors in, world model, policy, actuation. Everything outside that box gets filed under “operations” and handed to somebody else. Waymo’s growth curve suggests that boundary is drawn in the wrong place.
A deployed driving agent has hard dependencies that never appear in an architecture diagram. It needs permission to operate in a jurisdiction. It needs a legal definition of who holds the driving task. It needs insurance frameworks, curb access, airport pickup rules, and a crash-reporting regime it can actually satisfy. None of these are learned. All of them gate whether the policy network ever gets to run in a given ZIP code.
Scaling to a dozen new metros means clearing that dependency set a dozen times, against a dozen different city councils and at least several state legislatures. Meanwhile a competitor with a driver in the seat and an app already installed everywhere has no such gate. That asymmetry, not model quality, is what a lobbying budget is buying down.
Why generalization stops at the county line
There’s a genuinely interesting technical wrinkle here. We usually discuss generalization as a distribution-shift problem: new weather, new road geometry, new pedestrian behavior, unfamiliar signage. Dallas is not San Diego, and neither is Phoenix. Fair enough, and that work is hard.
But the shift that actually slows a rollout is often institutional. Two cities can have nearly identical driving conditions and wildly different approval timelines. Your model transfers fine; your operating license does not. That’s a form of non-stationarity no amount of additional data cures, because the environment includes agents that respond strategically to your presence: incumbent operators, labor groups, regulators with their own incentives.
Framed that way, the robotaxi contest looks less like a benchmark race and more like a multi-agent game where the other players get to modify the rules. Doubling spend on rule-modification is a rational move for whoever has the better car and the worse market access.
What this implies for how we build agents
A few things follow if you take the policy layer seriously as part of the stack:
- Constraint configuration becomes a first-class interface. An agent that can accept per-jurisdiction operating rules as data, rather than requiring a re-tuned policy per city, expands faster. City-specific behavior belongs in a declarative layer, not baked into weights.
- Auditability is a product feature, not overhead. If you cannot explain a specific decision to a specific regulator on a specific street, you have a capability gap regardless of your aggregate safety numbers. Logging and explanation pipelines are load-bearing.
- Evaluation should include the approval loop. We measure miles, interventions, and collision rates. Nobody publishes time-to-operational-authority per market, yet that’s the metric governing the growth curve.
- Operational surface counts. Depots, charging, cleaning, remote assistance, and roadside recovery all shape unit economics. Autonomy that requires heavy human tail support scales differently from autonomy that doesn’t.
The uncomfortable part
I find the lobbying escalation slightly deflating, and I think that reaction is worth examining rather than indulging. There’s a purist instinct in this field that the best system should win on merit, and that spending on political influence is what you do when the engineering isn’t finished.
That instinct is wrong here. Public road transport is a regulated commons. Any competent deployment of autonomous vehicles at city scale must involve negotiating with the public through its institutions. The alternative isn’t purity, it’s operating without consent. What deserves scrutiny is not that the negotiation happens but whether it’s transparent and whether safety claims made in legislative rooms match the evidence.
Waymo going from 250,000 to 450,000 weekly rides in about half a year is a real capability signal. Doubling down in Austin and lining up Dallas and San Diego is a real distribution signal. But the number I’d most like to see is not ride volume. It’s how much of the next 12 cities is unlocked by better software versus better lawyering, because that ratio tells you what kind of problem autonomy has actually become.
My working assumption is that the software is no longer the binding constraint. If that’s right, the most valuable engineering in the next two years will be the unglamorous kind: configuration systems, evidence pipelines, and interfaces that let a driving agent prove itself to a skeptical city in weeks instead of years.
đź•’ Published: