Agentic Commerce Interface (ACI) – Release Notes v1.0.0 (30 September 2026)
Overview
This release delivers the first production-capable version of the Agentic Commerce Interface (ACI) — a multi-tenant AI platform that lets commerce businesses deploy conversational AI agents that work directly with their commerce systems (Shopify, Salesforce, Equinox).
This release covers seven capability areas: Commerce Agent Orchestration, Discovery, Order & Cart, Subscriptions, Support, Connectors & MCP, and Organisation, Auth & RBAC.
Functional Changes – What’s New for Users
Commerce Agent Orchestration
Shoppers can now hold a single conversation that ACI automatically routes to the right specialist agent for each request.
- Supervisor + 4 specialist agents: A supervisor classifies each message and hands off to the Discovery, Order, Subscription, or Support agent. Each agent can be enabled or disabled per tenant.
- Dual agent runtime: Agents run on LangGraph (production default) or the Strands Agents SDK (alternate runtime), selectable per deployment.
- Live reasoning panel: The chat UI streams every stage of a turn — intent classification, retrieval, ranking, slotting, and tool calls — as it happens.
- Session continuity: Conversations persist per tenant, user, and session, so shoppers can pick up where they left off.
- Anonymous + authenticated access: Unauthenticated shoppers can use product discovery; order, subscription, and return actions require a verified identity.
Discovery
Shoppers can now find products through natural-language search that returns ranked, merchandised recommendations.
- Multi-stage retrieval: Three selectable engine versions (v1–v3) combine semantic knowledge-base search, hard-constraint filtering, and a 5-signal weighted ranking model.
- Slot-based merchandising: Results are composed into HERO / NUDGE / COMPLEMENT / BUNDLE slots rather than a flat list.
- Substitution & out-of-stock rerank: When a requested or subscribed item is unavailable, the agent surfaces the closest in-stock substitute with a transparent “why this” explanation.
- Consumer memory & personalisation: Long-term signals (concerns, preferred ingredients and brands, prior purchases) persist across sessions and influence future ranking.
- Seasonal & campaign awareness: Campaigns and seasonal boosts can be layered into ranking without code changes.
- Guardrailed output: Product image and link URLs are HTTPS-validated, results are capped at 20 products per component, and non-HTTPS or malformed media is dropped before reaching the client.
Order & Cart
Shoppers can now manage their cart, check out, and review past orders through conversation.
- Confirmation gate: An order can only be placed after a checkout-intent turn and an explicit affirmative reply in the same conversation, preventing accidental or injected order placement.
- Paginated order history: Capped at 25 orders per response.
- Live price and availability check: The Order agent re-fetches product details to confirm price and availability before checkout.
Subscriptions
Shoppers can now manage their subscriptions end to end through conversation.
- Subscription lifecycle: Create, view, pause, cancel, and change plan or frequency.
- Retention offers: Pause and cancel requests are evaluated for applicable retention offers before the action is completed.
Support
Shoppers can now get support answers and complete common post-purchase actions without contacting an agent.
- Two support engines per tenant: v1 (deterministic 6-stage playbook pipeline) and v3 (stateful workflow-graph engine with richer branching).
- Certified Equinox support actions: Order lookup, return label generation, tracking check, stock check, subscription lookup, and subscription cancellation.
- Document-grounded answers: Policy, FAQ, and product guide documents can be uploaded per tenant and are used to ground support responses, with a configurable similarity threshold.
- Escalation fallback: If no playbooks are configured for a tenant, support escalates instead of fabricating a resolution.
Connectors & MCP
Connectors let ACI agents work against external commerce platforms, and let external AI clients call ACI as a tool.
- Commerce adapters: Shopify, Salesforce, and Equinox connectors behind a single normalised interface (search, cart, order, and subscription operations), plus a local fixture adapter for demos and tests.
- Agent-to-Agent (A2A) protocol server: ACI is reachable by any A2A-compliant client, exposing one skill per specialist domain, authenticated via per-caller HS256 JWT.
- Hosted MCP endpoint: Any MCP-compatible client (e.g. Claude Desktop, MCP Inspector) can call ACI as a single remote tool (
ask_aci_commerce_agent) over a hosted Streamable HTTP endpoint — no local bridge script required. It reuses the per-caller secret already provisioned for A2A, so there is no separate API key to manage. - Equinox OAuth keep-alive: Equinox connector credentials are refreshed automatically so scheduled and event-driven agent runs don’t fail on expired tokens.
Organisation, Auth & RBAC
Platform administrators can now configure tenants, shopper authentication, and admin access control through the Admin console and API.
- Tenant model: Each tenant selects its commerce platform (Equinox, Shopify, Salesforce, or local) and is independently configurable for agents, support, discovery, and auth.
- Shopper-auth modes per tenant: Cognito-backed shopper login, or a storefront JWT-relay mode where each shopper’s own commerce credentials are used per session.
- Tenant JWT relay (RS256 / HS256 / ES256): Tenants can federate their own identity provider. ACI verifies the signed JWT against a published public key, with replay protection, issuer and lifetime checks, and strict algorithm pinning.
- RBAC: Role-based access control for admin operators with privilege-level checks, audit-logged role grants and revocations, and a dedicated super-admin bootstrap path.
- Admin MFA & lockout: TOTP MFA required by default for admin operators (configurable per deployment), escalating lockout on repeated failures, and CloudWatch alerting on anomalous login and MFA-failure volume.
- Password policy: Server-side enforcement of a 12+ character minimum with upper/lower/digit/symbol, rejection of passwords containing the user’s email, a blocked common-password list, and reuse prevention across the last 5 passwords.
API Changes
The following new and updated API endpoints are available in this release. All admin endpoints are served under /admin. Shopper chat is served over WebSocket; external agent and tool access is served over A2A and MCP.
Admin API — Tenant Management
| Method | Endpoint | What it does |
|---|---|---|
| GET / POST | /admin/tenants |
List or create tenants |
| GET / PUT / DELETE | /admin/tenants/:id |
Read, update, or delete a tenant |
| GET / POST | /admin/tenants/:id/users |
List or create tenant-scoped users |
| GET / PUT / DELETE | /admin/tenants/:id/users/:username |
Manage a single tenant user |
| GET / PUT | /admin/tenants/:id/auth-config |
Read or update shopper auth configuration |
| POST | /admin/tenants/:id/storefront-login |
Mint a per-shopper storefront session (JWT relay) |
| POST | /admin/tenants/:id/storefront-logout |
End a storefront session |
| GET / PUT | /admin/tenants/:id/agents / /agents/:agentId |
Configure per-tenant agent enablement and settings |
| POST | /admin/tenants/:id/jwt-secret |
Provision or rotate a tenant JWT relay secret |
| POST | /admin/tenants/:id/a2a/callers/:callerId/secret |
Generate or rotate an A2A caller secret (also used for MCP) |
| GET / PUT | /admin/tenants/:id/federation |
Manage identity federation settings |
| GET / POST / DELETE | /admin/tenants/:id/blocked-passwords |
Manage the tenant’s blocked-password list |
Admin API — Support Configuration
| Method | Endpoint | What it does |
|---|---|---|
| GET / POST | /admin/tenants/:id/support/playbooks |
List or create support playbooks |
| PUT / DELETE | /admin/tenants/:id/support/playbooks/:category |
Update or delete a playbook |
| GET / POST | /admin/tenants/:id/support/documents |
List or upload support documents |
| DELETE | /admin/tenants/:id/support/documents/:docId |
Delete a support document |
| GET | /admin/tenants/:id/support/policy-facts |
List extracted policy facts |
| PATCH | /admin/tenants/:id/support/policy-facts/:factId |
Edit a policy fact |
| PUT | /admin/tenants/:id/support/returns-config |
Update returns configuration |
| PUT | /admin/tenants/:id/support/ticket-settings |
Update ticket/escalation settings |
| PUT | /admin/tenants/:id/support/prompts/:stage |
Override the prompt for a support pipeline stage |
| GET | /admin/tenants/:id/workflows |
List support v3 workflow graphs |
| POST | /admin/tenants/:id/workflows/seed / /generate |
Seed default workflows or generate one |
| GET / PUT / DELETE | /admin/tenants/:id/workflows/:workflowId |
Read, update, or delete a workflow |
| POST | /admin/tenants/:id/workflows/:workflowId/validate |
Validate a workflow graph |
Admin API — Discovery, Order & Subscription Configuration
| Method | Endpoint | What it does |
|---|---|---|
| GET / PUT / POST / DELETE / PATCH | /admin/tenants/:id/discovery/:resource / :resourceId |
Manage discovery configuration (ranking weights, seasons, campaigns) |
| GET / PUT / POST / DELETE | /admin/tenants/:id/discovery/slots |
Manage merchandising slots |
| GET / PUT / POST / DELETE | /admin/tenants/:id/discovery/guardrails / /guardrails/soft |
Manage hard and soft discovery guardrails |
| GET / PUT / POST / DELETE | /admin/tenants/:id/discovery/concerns / :resourceId |
Manage shopper concern taxonomy |
| GET / PUT / POST / DELETE | /admin/tenants/:id/discovery/catalog-import / :resourceId |
Import and manage product catalogues |
| GET / POST | /admin/tenants/:id/discovery/pips / presign / queue / :resourceId/download / :resourceId/requeue |
Upload, ingest, download, and re-queue product information documents |
| GET | /admin/tenants/:id/order / /subscription |
Read order or subscription domain configuration |
| PUT | /admin/tenants/:id/order/:resource / :resourceId |
Update order domain configuration |
| PUT | /admin/tenants/:id/subscription/:resource / :resourceId |
Update subscription domain configuration |
Admin API — Consumer Memory
| Method | Endpoint | What it does |
|---|---|---|
| GET | /admin/tenants/:id/memory/sessions |
List stored conversation sessions |
| GET / DELETE | /admin/tenants/:id/memory/sessions/:sessionId |
Read or delete a session |
| GET | /admin/tenants/:id/memory/consumers |
List consumers with stored memory |
| GET / DELETE | /admin/tenants/:id/memory/consumers/:userId |
Read or erase a consumer’s memory |
| GET | /admin/tenants/:id/memory/consumers/:userId/narrative |
Read a consumer’s memory narrative |
| POST | /admin/tenants/:id/memory/consumers/:userId/entries |
Add a memory entry |
| PUT / DELETE | /admin/tenants/:id/memory/consumers/:userId/entries/:entryId |
Update or delete a memory entry |
Admin API — Platform & RBAC
| Method | Endpoint | What it does |
|---|---|---|
| GET | /admin/agentcore/endpoints |
List available Bedrock AgentCore runtime endpoints |
| GET | /admin/bedrock/models |
List available Bedrock model IDs |
| POST | /admin/shopify/register / /admin/shopify/uninstall |
Shopify app registration lifecycle |
| GET / POST | /admin/users |
List admin operators or invite a new one |
| GET | /admin/rbac/privileges / /admin/rbac/privilege-sets |
List available privileges and privilege sets |
| GET / POST | /admin/rbac/roles |
List or create roles |
| GET / PUT / DELETE | /admin/rbac/roles/:roleId |
Manage a single role |
| POST / DELETE / GET | /admin/rbac/users/:sub/roles / /roles/:roleId |
Grant, revoke, or list a user’s role assignments |
| GET | /admin/rbac/users/:sub/effective-privileges |
Resolve a user’s effective privileges |
| GET | /admin/rbac/audit |
Paginated RBAC audit log |
Agent Runtime Endpoints
| Protocol | Endpoint | What it does |
|---|---|---|
| WebSocket | wss://.../ws?tenantId={id} |
Primary shopper chat endpoint (chunked frames). The token is carried in an HttpOnly cookie (or in the message body for the cross-site embedded widget) and re-validated on every frame — never in the URL |
| HTTPS (A2A) | https://.../a2a (message/send, .well-known/agent-card.json) |
Agent-to-Agent protocol endpoint, HS256-authenticated |
| HTTPS (MCP) | https://.../mcp/:tenantId/:callerId |
Hosted MCP endpoint exposing ask_aci_commerce_agent, Bearer-token authenticated against the tenant’s A2A caller secret |
Technical Stack
Frontend
| Component | Technology |
|---|---|
| Framework | React 19 + Vite 8 |
| UI library | MUI 7 + Infosys Equinox AIUX design system + framer-motion 12 |
| Routing | React Router 7 |
| Realtime | Native WebSocket client with chunked-frame reassembly |
| Architecture | Single-page app containing the shopper chat UI and the Admin console (tenant, RBAC, support, discovery, order, and subscription configuration) |
Backend
| Component | Technology |
|---|---|
| Language | TypeScript (ESM), Node.js 22 / 24 |
| Agent runtime | LangGraph 1.x (production default) + Strands Agents SDK (alternate runtime) |
| LLM model (default) | Claude Sonnet 4.6 (primary) and Claude Haiku 4.5 (fast paths) via Amazon Bedrock |
| AI guardrails | Prompt-injection detection on all user input, PII/PCI redaction on client-facing output, per-agent action policy caps, contextual grounding checks |
| Knowledge base | Amazon Bedrock Knowledge Bases + OpenSearch Serverless (vector search, Titan Embed Text v2) |
| Compute | Amazon Bedrock AgentCore (container runtime) + AWS Lambda (Admin API, A2A, MCP, catalogue and support ingest) |
| Database | Amazon DynamoDB (tenants, sessions, configuration, playbooks, policy facts, RBAC) |
| File storage | Amazon S3 (catalogues, knowledge-base documents, support documents, UI assets) |
| Secrets / credentials | AWS Secrets Manager (per-tenant JWT secrets, A2A/MCP caller secrets, commerce credentials) |
| Authentication | Amazon Cognito (per-tenant shopper pools + separate admin pool) + tenant JWT relay (RS256 / HS256 / ES256) |
| Connector protocols | MCP (Streamable HTTP), REST |
| Observability | OpenTelemetry auto-instrumentation + Amazon CloudWatch |
Deployment
| Component | Technology |
|---|---|
| Infrastructure as code | AWS CDK v2 (TypeScript), stacks: network, database, agentcore, bedrock-agent, gateway, websocket-proxy, admin, a2a, mcp, catalog, search, storage, support, secrets, observability, pip, ui, shopify-app/host |
| Container | Docker (Node.js, multi-stage build) for the AgentCore agent runtime |
| CI/CD | GitHub Actions (PR checks: lint, dependency-cruiser, Jest; deploy: build, test, cdk deploy --all on merge to master) |
| CDN / Edge | Amazon CloudFront with Origin Access Control and path-based routing per service (/a2a, /mcp/*) |
| Testing | Jest (unit) + Playwright (end-to-end) |
Upgrade Notes
- WebSocket auth: The session token is re-validated on every frame, not just at connection time. A frame carrying an expired or invalid token is rejected and the client must re-authenticate before sending further messages.
- A2A / MCP callers: Caller secrets are provisioned via
POST /admin/tenants/:id/a2a/callers/:callerId/secretand are reused as-is for MCP access — there is no separate MCP API key endpoint. - Support engine selection: Tenants on the playbook pipeline (
supportEngine: 'v1') are unaffected. Switching a tenant to'v3'requires configuring workflow graphs via/admin/tenants/:id/workflowsbefore support will resolve anything beyond escalation. - Shopper auth mode: The default mode (
shopperAuthMode: 'cognito') uses a shared commerce connector account. For per-shopper data isolation, configurestorefrontmode with tenant JWT auth. - Order placement confirmation gate: Any integration driving the Order agent programmatically must send an explicit checkout-intent turn followed by an explicit affirmative turn — a single-shot “place order” message is rejected with
confirmation_required. - Result caps: Clients consuming order history or discovery results must handle pagination and truncation — responses are capped at 25 orders and 20 products per component.
- Wallet / Credits: Not available in this release. Wallet or store-credit queries return a “not yet available” message.
Revision History
2026-10-05 | AN – Created the page and added the content.