Car electrification was never just a new engine. It's an infrastructure problem: without a charging network, without standards, without battery recycling, the electric vehicle stays a niche object—no matter how good the motor is.
I was reflecting on how difficult it is to move from one mode of software engineering to another—why the transition resists, why it drags, why even teams that want to change find themselves stuck halfway. That's when the analogy struck me.
The harness is the charging network
We tend to treat the agentic "harness"—the environment that frames, contextualizes, and verifies an agent's work—as just another productivity tool. In reality, it's the equivalent of the charging network: without it, electrification (read: agent autonomy) isn't viable, no matter how powerful the model behind it is. And like charging stations, it doesn't get built overnight. It takes virtuous loops—logs, evals, feedback, adjustments—that look a lot more like infrastructure under construction than a product you buy off the shelf.
The gearbox problem
But the deeper shift isn't about the motor—it's about the gearbox. Combustion engines produce power only within a narrow RPM range; the gearbox exists to compensate for that constraint. Software organizations have built their own transmission: tickets, sprint ceremonies, detailed specifications, sequential handoffs. These aren't intrinsically valuable. They exist because humans are relatively slow implementation engines, and the entire delivery process was designed around that limitation. Agentic systems may not need as much of that transmission. The gearbox doesn't disappear overnight—but its necessity decreases, and with it, much of what we've come to think of as "software engineering process."
An electric motor doesn't just replace the engine—it changes the physics. The torque profile, the RPM range, the instant power delivery: all of it demands that the rest of the vehicle be re-engineered around the new source of energy, not just bolted onto it. (Which is also why EVs wear tyres faster—a fact I find worth mentioning here.) Agentic workflows impose the same pressure on software delivery: different pace, different failure modes, different parts wearing out—code review, governance, comprehension debt, trust—not the code itself.
The weight of transition
The financial parallel is the most uncomfortable one. The cost per mile eventually drops—but retiring the existing fleet carries its own weight: depreciation, running contracts, retraining. And there will be dumping: players selling agentic work at a loss to grab market share before the infrastructure is ready to support it.
Hybrid may not be just a phase. Just as EVs thrive in dense, well-wired urban fabric and struggle elsewhere, AI-operated engineering will shine on greenfield projects with the right harness in place, and stay partly hybrid wherever the brownfield context resists it. The psychological barrier isn't capability—most teams can see what these systems can do. It's confidence: whether the system can be trusted enough to remove the process layers that exist to compensate for human limitations. That may keep the hybrid era longer than enthusiasts expect—and more permanent than they'd like.
The sustainability question
Electric vehicles are only as clean as their supply chain. The lithium has to be mined, the electricity generated, the batteries eventually recycled. The environmental virtue of the drivetrain doesn't survive an unsustainable production chain. Agentic engineering faces an equivalent question. The models were trained on decades of human-generated code, writing, and reasoning. The open-source ecosystems that fed them were built by engineers who understood what they were building. If agentic adoption erodes the formation pipeline—producing teams that can direct agents but no longer understand what they produce, or exhausting the intellectual commons the next generation of models will need—the efficiency gains are real but the system isn't sustainable. Comprehension debt is the carbon footprint of agentic engineering. It doesn't show up in the sprint velocity metrics.
In 1899, on Michelin tyres, Camille Jenatzy's La Jamais Contente—a torpedo-shaped electric vehicle—became the first to exceed 100 km/h. The engine was never in doubt. What took a century was making electric mobility viable at scale—the charging network, the standards, the infrastructure that turned a record into a mode of transport. Building that equivalent today—harness, feedback loops, engineering culture—is the real challenge facing every product team that wants to turn this new power into something durable.
