Picture a fourth-floor conference room in a state capitol building. The carpet is that particular shade of institutional blue. Across the table, a legislative aide is asking a question that no perception stack can answer: who is legally the driver when there is no driver? On one side of the table sits someone representing Waymo. Somewhere down the hall, someone representing Uber is asking for a meeting on the same bill.
That room, and hundreds like it, is where a meaningful slice of the robotaxi outcome gets decided. Waymo has doubled its lobbying spending as tensions with Uber deepen. Uber has pledged $10bn to win the robotaxi race, a strategy shift for a company that once outsourced its autonomy ambitions. Waymo, meanwhile, finalised a $16bn funding round at a $110bn valuation and is reportedly exploring a split from Uber altogether.
As someone who spends most of her time thinking about agent architecture, I find the lobbying line item more instructive than the funding line item. It tells you something specific about where the hard problems have migrated.
The Stack Nobody Draws in the Architecture Diagram
When we sketch an autonomous agent, we draw sensors, then perception, then prediction, then planning, then control. Sometimes we add a fleet orchestration layer on top. What we almost never draw is the layer above that: the regulatory interface. Yet for any agent that acts in the physical world at scale, that layer is load-bearing.
Consider what a permit actually is from a systems perspective. It is a constraint on the agent’s action space, defined by an external authority, revisable without notice, and enforced with penalties that dwarf any internal loss function. An operational design domain is not just an engineering boundary. It is a negotiated boundary. The set of streets a vehicle may drive, the hours it may operate, the conditions under which it must hand off or stop, the reporting obligations after an incident — these are parameters set in conference rooms, not in training runs.
So doubling lobbying spend is not a distraction from the technical work. It is an investment in widening the feasible region that the technical work operates inside. You can have the best planner in the world and still be confined to a geofence someone else drew.
Two Different Bets on Where Intelligence Lives
The Waymo and Uber positions encode genuinely different architectural theses, and the possible split makes that clearer.
Waymo’s approach has always been vertically integrated: own the sensing, own the driving policy, own the vehicle platform, own the safety case. In that model, the demand layer — the app, the rider matching, the surge pricing — is a commodity you can rent. Uber was a distribution channel. A $110bn valuation and $16bn in fresh capital changes the calculus on renting versus building that channel yourself.
Uber’s $10bn commitment reads as the mirror image. Uber owns the demand graph, the rider habits, the routing history across millions of trips. Its bet has been that the driving agent eventually commoditises — that many suppliers will produce adequate autonomy, and the durable advantage sits in matching supply to demand at scale. Committing $10bn to robotaxis directly suggests less confidence that suppliers will stay plentiful and cooperative.
Both are making claims about which layer of the stack captures value. That is a familiar argument in software, and it usually resolves messily.
What This Means for Anyone Building Deployed Agents
The lesson generalises well beyond vehicles. Any agent that takes consequential action in a regulated domain — clinical workflows, financial transactions, hiring, physical logistics — will eventually meet the same structure. The capability frontier and the permission frontier advance at different speeds, and the second one moves through institutions rather than compute.
A few implications I would take seriously:
- Design your agents so the constraint set is explicit, inspectable, and swappable. If your operating boundaries are baked into learned weights rather than declared as policy, you cannot respond to a rule change without retraining.
- Treat auditability as a first-class requirement, not instrumentation you bolt on later. The artifacts regulators ask for are the artifacts that make debugging possible anyway.
- Assume your distribution partner may become your competitor. Architectural dependencies on a partner’s demand layer are strategic dependencies.
- Budget for the regulatory interface the way you budget for infrastructure. It is a real cost of operating a physical-world agent, not overhead.
The romantic version of autonomy is a solved driving policy, generalising cleanly across cities. The actual version is a solid driving policy negotiating city by city, with lawyers in the loop and a fleet ops team watching dashboards. Waymo’s lobbying budget and Uber’s capital pledge are both admissions that the remaining distance is not purely a modelling problem.
For those of us designing agent systems, that is worth internalising early. The environment your agent operates in has stakeholders, and they get a vote.
🕒 Published: