A resident asks the city a simple question: "Can I get a permit to close my street for a block party, and will it affect the bus route?" Today that question touches three departments, three systems, and often three phone calls. As cities add AI assistants, the risk is that each department builds its own, with its own data store, security review, and audit trail, and none of them can pass the question to another.

This talk shares an architecture called MicroAgents, drawn from running AI agents in production for large enterprises, and applies it to city services. Each department owns only two things: a tool server that exposes its own data and actions, and a prompt that holds its domain knowledge. A shared city platform owns everything else, including identity, audit logging, tracing, and memory.

The talk covers how one agent hands a request to another while keeping the full context, how shared memory lets a rule learned by one department (such as "trash pickup moves one day later in a week with a city holiday") help every other department, and how memory can be scoped by resident, department, and city. It also covers why full tracing of every agent step matters for public accountability, and which parts of the platform are worth buying instead of building.

The goal is to help cities add new AI services in weeks, without rebuilding the same foundation for each department.