At Lund University I started a project that became Uniti, an attempt to rethink the electric city car. We attracted talent, investors, industrial partners, press and thousands of believers. We also met the merciless reality of turning an ambitious prototype into a manufactured product.
Physical systems do not care about the quality of your pitch. Components have lead times. Weight moves somewhere else. Certification is a gate, not a vibe. A beautiful idea survives only when thousands of constraints agree to let it exist.
Agents also live in the real world
Software agents feel weightless, which makes it easy to mistake a convincing demo for an operating system. But the moment an agent touches a customer, sends a message, changes a record or makes a financial decision, it enters a world of consequences.
The useful question is not ‘Can the model do this?’ It is ‘Can the system do this repeatedly, under partial information, while exposing failure clearly and recovering safely?’ That is an engineering question.
Design the failure before the magic
In hardware, you model loads and tolerances. In agentic systems, the equivalent is permission scope, confidence thresholds, idempotency, human review and observability. What happens when an action runs twice? When the source is stale? When the model sounds certain and is wrong?
A serious agent is not a personality wrapped around a prompt. It is a bounded actor inside a designed operating environment.
Ambition needs machinery
Uniti taught me that attention, belief and vision can create extraordinary momentum, but momentum is not production. Today I apply that lesson to software: start with the outcome, trace every dependency, automate the repeatable path and make exceptions visible.
AI expands what a small team—or one person—can attempt. The founders who win will combine that ambition with a respect for constraints. The machine has to work after the room stops applauding.