Request and data flow

Regardless of the entry channel, every AVX interaction follows the same high-level flow.

Channel ingress

A request enters AVX through one of four channels:

• Voice (Amazon Connect + Nova 2 Sonic) — the caller’s speech is streamed to the AgentCore bidi runtime. Amazon Nova 2 Sonic performs speech-to-speech transcription and synthesis within the runtime container. The supervisor receives the transcribed text; its response is synthesized back to speech before being streamed to the caller.

• Chat (Amazon Connect Chat) — text messages from the Connect Chat widget arrive at the supervisor via a Lambda integration. The response is returned as text.

• Web / mobile — a POST to the REST API endpoint {vsaApiEndpoint}/assistant/chat carries the user prompt, session ID, and VIN. The Cognito authorizer validates the caller’s JWT and injects verified claims into the Lambda context; the Lambda proxy infers the persona server-side from cognito:groups and custom:role claims, and forwards to the text AgentCore runtime.

Supervisor orchestration

Once the supervisor receives an utterance, it executes the following sequence:

1. Session context load (first turn only) — persona_loader pre-fetches ADP tabular data via cross-region Athena and seeds the result into AgentCore Memory LTM. On subsequent turns within the session, all persona context is read from Memory without Athena round trips.

2. Live state read — for vehicle-linked sessions, the supervisor reads the Redis LKS snapshot from CMS. This snapshot includes active DTCs, telemetry readings, and open-recall flags. The snapshot is attached to the session context and passed as structured arguments to tools that need it.

3. Intent classification and triage — for any message that references vehicle symptoms, warning lights, or service urgency, the supervisor calls triage() first. The triage result constrains the rest of the turn: a P0 result delivers the stop-driving instruction word for word and offers the driver a human agent and roadside assistance, each one tap; P1 and P2 results shape the urgency framing of the booking recommendation.

4. Domain tool dispatch — based on intent and triage result, the supervisor dispatches to the relevant domain tools (vehicle_identity(), knowledge(), parts_lookup(), service_network()). Tool calls are sequential by default and parallel where dependencies allow; Strands' native tool-call streaming returns partial results as they arrive.

5. Booking or escalation — if the interaction reaches a booking decision point, the supervisor calls book(). If it reaches an escalation trigger (P0, customer request, policy exception), it calls escalate().

6. Response synthesis — the supervisor synthesizes a grounded response from tool results. For voice, Nova 2 Sonic streams audio back to the caller. For text/web/chat, a JSON response carries the assistant text, triage classification, VIN, and citation references.

Grounding data path

AVX’s two grounding surfaces are architecturally independent:

• Tabular data (AgentCore Memory LTM) — ADP Athena queries → session-init LTM seed → per-turn metadata-filtered Memory Retrieve. Latency: ~6–7 seconds for session-init; sub-100ms for per-turn Memory reads.

• Reference data (Bedrock KB Retrieve) — cross-account resource policy → Bedrock Retrieve with source-category filter → top-k passage retrieval from the Amazon S3 Vectors index. Latency: 200–600ms per Retrieve call, depending on index load and query complexity.

If Memory is unavailable or returns no records, tools return structured unavailable messages; the session degrades gracefully without a hard failure. If KB Retrieve fails, the knowledge() tool returns an error trace and the supervisor informs the customer that reference lookup is temporarily unavailable.

Tier 2 trigger and artifact flow

Tier 2 agent flows run independently from the Tier 1 request-response path. An EventBridge Scheduler rule or event rule fires the trigger; a Lambda invokes the agent module; the agent queries ADP Athena tables or DMS APIs; the result is written to vsa-{stage}-avx-findings; the DynamoDB stream fires the outreach broker. No customer session is open during this flow.

The Tier 1 flow intersects the Tier 2 platform at two points:

1. Finding read tools — get_energy_advice(), get_health_findings(), get_coverage(), get_loyalty_standing(), and get_ownership_position() retrieve the most recent Finding for the current session’s scope from vsa-{stage}-avx-findings via DynamoDB point reads. The supervisor calls these tools the same way it calls any other domain tool; the result is a pre-computed conclusion with evidence attached, not a live reasoning call.

2. Proactive session entry — when the broker delivers a push notification and the owner taps it, the mobile client issues a POST /assistant/chat with avx_finding_id in the body. The supervisor loads the referenced Finding’s evidence into session context before the first response, so the conversation opens at the Finding’s depth.