Customer support app

The customer support example shows a smaller workflow app. It is intentionally less complete than the Content Factory fixture, but it demonstrates how Agent App applies outside content production and still composes Runtime, UI, Context, Knowledge, Skills, Tools / Connectors, Artifacts, Evidence, Policy, and QC.

Reference package: docs/examples/customer-support-app/APP.md

User job

A support team wants grounded replies and escalation notes that follow product facts and support policy. The app should help agents answer faster without inventing policy or leaking customer data.

Package shape

The example declares:

• Runtime intent for structured reply drafting
• product and policy Knowledge templates
• a draft_reply command entry
• optional ticket lookup Tool / Connector requirement
• reply_draft Artifact type
• policy_compliance Eval
• app.runtime.yaml for structured reply drafts, approval, session, tool discovery, checkpoints, and observability
• support category presentation metadata

It is a draft app, so validation still warns that a product-level runtime package, full layered config, and explicit permissions should be added before production use.

Suggested complete version

A production support app would usually add:

Area	Additions
Entries	Reply drafting, policy lookup, escalation note, manager review.
UI	Ticket side panel and policy citation panel.
Storage	Draft history, escalation state, policy version snapshots.
Tools	Ticket lookup, CRM update, customer profile, macro export.
Artifacts	Reply draft, escalation note, policy citation bundle.
Evals	Policy compliance, tone, source grounding, privacy check.
Permissions	Ticket read, optional ticket write, CRM tool execution, and runtime approval for risky actions.
Secrets	CRM OAuth or tenant ticketing connector handle.

Boundary example

Asset	Correct place
How to write an empathetic reply	Agent Skill
Product facts and refund policy	Agent Knowledge
Ticket lookup connector	Agent Tool
Draft reply command and review workflow	Agent App
Reply draft with JSON Schema	Agent Artifact
Policy compliance result	Eval and Evidence

Why this matters

Support workflows are high trust. The app should never answer from generic model knowledge when policy Knowledge is required. Readiness should block or warn when support_policy is not bound.

Try it locally

npm run cli -- validate docs/examples/customer-support-app
npm run cli -- project docs/examples/customer-support-app
npm run cli -- readiness docs/examples/customer-support-app

The draft fixture is useful for testing warnings, runtime contract projection, and progressive completeness.

End-to-end support flow

Open ticket
→ retrieve product facts and support policy
→ draft answer with citations
→ run policy and privacy checks
→ human support agent edits or approves
→ create reply draft artifact
→ optionally export to ticketing system

The app should keep the human support agent in control. It can accelerate retrieval, drafting, summarization, and escalation note creation, but final customer communication should remain reviewable and traceable.

Entry design

Entry	Kind	Purpose
draft_reply	command	Draft a response for the active ticket or pasted customer message.
policy_lookup	panel	Show relevant policy snippets with source links.
escalation_note	workflow	Build an internal note when the case requires specialist review.
reply_history	artifact	Review durable reply drafts and policy evidence.
support_settings	settings	Bind policy Knowledge, tone rules, and ticket connector permissions.

Safety expectations

• Required policy Knowledge must be bound before answers can be marked ready.
• PII should stay in host-controlled workspace storage and Evidence, not in the public package.
• Tool calls that write back to CRM or ticketing systems require explicit permission.
• Low-confidence replies should be saved as drafts, not sent automatically.
• Readiness should surface missing product facts, stale policy versions, and missing connector auth separately.

How to adapt it

For a smaller team, keep only draft_reply, product facts, support policy, and manual export. For a regulated support team, add escalation workflow, manager review, redaction, audit exports, and stricter Evals. The same Agent App contract supports both without changing host core.