Bedrock + AppSync: A GraphQL Front Door for Agentic Workflows

Agentic workflows don't complete in one request/response cycle. A multi-step Bedrock agent call can take anywhere from a couple seconds to well over a minute depending on how many tools it invokes, and a frontend built around a synchronous REST call handles that badly — either a request that times out, or a client stuck polling an endpoint every few seconds hoping something changed.

Why AppSync fit

AppSync's subscriptions gave us a way to treat a long-running agent invocation like an async job with a live status feed, without hand-rolling WebSocket infrastructure. The flow: a mutation kicks off the agent run and returns immediately with a run ID; the client subscribes to that run ID; and as the agent progresses through tool calls, each step publishes an update the subscription pushes straight to the UI.

The type safety was the part I didn't expect to matter as much as it did. A GraphQL schema forces you to actually define the shape of an agent's intermediate state — what a "tool call in progress" event looks like versus a "final answer" event — instead of letting that shape drift implicitly across a REST API's ad hoc JSON payloads.

What this replaced

• Polling — no more client-side interval hitting a status endpoint; the subscription pushes only on real state changes.
• Timeout-prone long-lived HTTP requests — the mutation returns fast, the actual work happens async behind it.
• Untyped intermediate state — every event on the subscription is a defined GraphQL type, not a loosely-shaped blob the frontend has to defensively parse.

The trade-off

AppSync subscriptions add infrastructure surface — resolvers, a pub/sub layer, IAM scoping on who can subscribe to which run — that a simple REST endpoint doesn't need. For a single quick model call, that's overkill. For a multi-step agent workflow where the frontend genuinely needs to show live progress, it removed an entire category of "did the request actually finish" bugs.