Failure is the product now.
That inversion is what makes Laplace Intelligence interesting to me as someone who spends most of her time staring at agent architectures. The University of Michigan spinout, based in Ann Arbor, uses AI agents and interactive simulation to stress-test robots and self-driving vehicles before they get deployed. It is automated red-teaming for physical AI — deliberately trying to make a system fail, at scale, before a pedestrian or a warehouse worker becomes part of the test set. The company is competing in TechCrunch Disrupt’s Startup Battlefield, and notably, it is already profitable.
Why adversarial agents beat scenario libraries
Most autonomy validation I have seen over the past decade works like a checklist. Engineers enumerate scenarios — the unprotected left turn, the occluded cyclist, the stopped vehicle in a tunnel — and replay them with parameter sweeps. That approach has a structural ceiling: you can only test the failures you were clever enough to imagine. The long tail of physical-world behavior is not a list. It is a distribution with no clean boundary, and your scenario library samples whatever your team’s collective imagination happened to cover.
Putting agents in the adversary seat changes the search problem. Instead of enumerating cases, you give an optimizer an objective — induce a violation, a collision, a dropped payload, a planner timeout — and let it explore the space of perturbations. The interesting part is not that this finds more bugs. It is that the bugs it finds are qualitatively different. Adversarial search tends to land on the geometry of the failure surface rather than individual points on it: the specific combination of speed, occlusion timing, and actor intent where a policy’s confidence quietly inverts. That is the kind of finding that actually changes a controller, not just a test report.
The fidelity tax
There is an obvious tension here, and it is the one I would press hardest on. Adversarial search is only as trustworthy as the simulator underneath it. Push an optimizer against a high-fidelity environment and it will exploit the simulator as happily as it exploits the policy. Every robotics team that has run aggressive domain randomization knows this feeling: you fix a failure that only ever existed in your physics engine’s contact model.
So the technical bet underneath Laplace’s pitch is really about calibration. High-fidelity simulation plus agentic red-teaming is a reasonable architecture, but the hard engineering sits in deciding which discovered failures transfer to reality and which are artifacts. Done well, you get a filter that ranks findings by real-world plausibility. Done poorly, you get a firehose of unreproducible edge cases and an engineering team that stops reading your reports. That discrimination layer is where I would expect the durable intellectual property to live, more than in the adversarial search itself.
Partnerships as a grounding signal
The collaborations are telling. Laplace is working with partners including T-Mobile and Nvidia on smart city initiatives and warehouse robotics. Those two domains are not arbitrary picks. Warehouse robotics gives you a constrained, heavily instrumented environment where simulated predictions can be checked against logged reality at volume — exactly the feedback loop you need to calibrate a simulator. Smart city work pushes in the opposite direction, toward multi-agent traffic where the behavior of other actors is the dominant source of uncertainty and where an adversarial agent has the most room to be creative.
Running both in parallel is a sensible way to build out a validation stack. One side keeps you honest about physics. The other keeps you honest about intent modeling.
The business problem is the technical problem
Laplace is working to turn project-based contracts into recurring revenue, and I would argue that transition is not just a go-to-market question. Project work means a human expert configures the red-teaming campaign for each client, interprets the results, and hands over a report. Recurring revenue means the loop runs continuously against a customer’s evolving policy stack with far less hand-holding — moving the human further out of the loop, which is also the explicit direction of the product.
Getting there requires the adversarial agents to generalize across robot morphologies, sensor suites, and operational design domains without being re-tuned from scratch each time. That is the same generalization problem the rest of us are chasing in agent research, just pointed at testing instead of task completion. A profitable services business funding that work is a reasonable position to be in.
What I find most encouraging is the framing. Treating safety validation as an adversarial search problem rather than a compliance exercise is the right instinct for physical AI. The systems we are deploying are too behaviorally complex for checklists, and the cost of discovering that in production is measured in something other than engineering hours.
🕒 Published: