Divide a multiplayer netcode build with an agent

You are building real-time multiplayer for a small browser game in TypeScript. Clients send inputs, and the server runs the authoritative simulation at a fixed tick rate. Each client predicts its own player's movement at once, then corrects itself when server snapshots arrive. You want an agent to do much of the work.

How do you divide the build? Decide what the agent writes, what you own, and where you check before moving on.