Latest in AI
Amazon Connect Live Sync: Add One Screen Action Without Giving the Agent the Whole Page
A practical way to scope Touchpoint context, validate Live Sync commands and keep the final submission with the customer.
A support agent can help a customer select a delivery window without being responsible for changing the order. That's a useful boundary for your first Amazon Connect Live Sync integration: let the conversation update a visible choice, but make the customer confirm through the normal application flow.
AWS's November 20 guide distinguishes modalities, which render structured UI in the conversation panel, from Live Sync, which can drive the host page. Live Sync itself became generally available in September. The November walkthrough is useful because it explains what your frontend actually has to implement.
Check the workspace before wiring the frontend
Start with the Connect Customer instance and check regional availability for Agentic CX Designer and its flow block. The designer uses separate workspaces. A user needs the right role in the workspace containing the application; access in one workspace does not carry over to another.
Use a test workspace and a deployed application build. The Connect flow invokes it through an Agentic CX block, and Touchpoint associates the page with the active contact. This example uses the web SDK; AWS identifies the Android SDK as developer preview.
When an action is missing during testing, check the instance, workspace, deployed build and contact before changing the model prompt. If the page is connected to the wrong interaction, more prompt instructions will not fix it.
Send the options, not the whole form
Automatic frontend context is enabled by default. It can collect form controls, values and navigation links. For a delivery-window pilot, set automaticContext: false in the Touchpoint liveSync configuration and construct a narrow context update using the Live Sync node and Context API documentation.
Send only the data needed for this turn: the selection screen's state, the visible slot IDs and their display labels. Do not send a full order object, even if it is already present in the browser. The agent does not need an address or a payment field to highlight Tuesday afternoon.
Turning automatic context off does not make the exchange private or local. Values you send manually still go to the active Live Sync conversation. Review the payload itself and document which fields leave the frontend, regardless of how the SDK collected them.
Make the handler reject stale choices
Declare one action on the Live Sync node, such as select_delivery_window. Give it a scope tag for the selection screen, an input that represents the available options, and an output that carries a slot ID. Register a frontend handler for that same action name.
In the web integration, AWS warns that the declared output schema does not validate the returned value for you. Check both its shape and whether the ID is present in the currently displayed list. Validate against current state when the command arrives, not against the list that was visible when the conversation began.
Consider a customer who selects a slot as it sells out. If availability changed between the spoken request and the command, keep the previous visible state or ask the customer to choose again. Do not select a different slot silently.
Keep “selected” separate from “saved”
Give the handler one job: update a reversible draft selection. Do not let it call the order-update endpoint. The customer's confirmation should use your usual application route, which should check authorization and availability again before saving.
The conversation should distinguish the two steps as well. A response that says a slot is highlighted is different from one that says the delivery has been changed. Reserve the latter for a successful application response, not the model's inference about what happened.
Test the journey without the screen
Run four cases before expanding the rollout: an available slot, a stale slot, a command after the customer leaves the selection page, and a call with no page connected. Inspect the actual context sent in the test and have a frontend and contact-center reviewer agree on what each outcome should look and sound like.
For callers without a connected page, offer a spoken choice or hand off to a supported path. Screen actions should save people a step, not make the service harder to use when the screen is unavailable.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help AI teams explain technical products through clear documentation, practical guides and developer-focused content.