The empire you agree to dissolve

For years the right answer was custom. To run the same Assistant logic on servers and on a zoo of heterogeneous devices, my org owned a portability framework that crossed languages and surfaces — the kind of bespoke infrastructure you build when nothing off the shelf fits. It worked. It was also exactly the sort of thing that turns into a permanent tax: a custom solution that every new engineer has to learn, that only your team can maintain, that quietly argues for its own continuation.

When Assistant’s strategy tightened toward the Android ecosystem and the company’s attention swung toward the LLM era, I had to write a charter that pointed away from the work my own people did best. The principle we put at the top was deliberately uncomfortable: bias hard toward reusing standard, widely supported technologies, and stop creating Assistant-specific ones. For every capability we owned, three honest options — chart a path onto core infrastructure, retire it, or invest in it on purpose. ‘Continue because we already built it’ was not on the list.

That triage is easy to say and expensive to mean. Charting a path to core meant migrating onto a portable language we didn’t historically use and moving staff off systems they were proud of. Retiring meant admitting that something competent should still die. The case I’m least conflicted about, in hindsight, was the least glamorous: divesting from legacy on-device smart-home support that carried a standing maintenance burden of four or five engineers forever. Killing it didn’t ship a feature. It returned people to work that the future needed.

A platform that only your team can maintain isn’t leverage. It’s a debt you’re paying in other people’s headcount.

Then the harder version arrived. When the business pulled back its investment in on-device, the responsible move was to wind down teams I had personally stood up. I could have fought to keep them — there’s always a deck that justifies it. Instead I negotiated each system to a long-term owner, found new homes for the people who wanted them, and redirected the rest toward the quality and reliability work that mattered more. None of that shows up as a launch. All of it is the job.

What I keep turning over is the cost of doing it well. Dismantling cleanly looks, from a distance, identical to having less to defend. The engineer whose system you retired with care and the engineer whose system simply lost funding can end up in the same place, and only one of them was led there. I told my teams the truth: the strategy changed, the work is still respected, here is where you land. And I still don’t know if honesty during a wind-down reads as leadership or just as the softer face of a decision made above me.