Open the landing page. Explain that a production support Agent needs more than a chat box: it must retrieve policy, inspect live business state, expose safe execution evidence, and stop before sending risky answers.
- Open Knowledge and upload
seed/knowledge/subscription-activation.md. - Open Customer Portal.
- Ask:
My Pro payment succeeded but workspace is still Free. - Point out the visible stages, three read-only tool calls, cited policy excerpt, and generated answer.
- Show that the P1 guard does not send the draft. Create a Ticket.
- Open Agent Workspace.
- Select the P1 Ticket.
- Show intent, confidence, citations, and business facts.
- Click 批准并发送 and explain the confirmation gate.
- Confirm approval and show 已解决.
- Mention that review, final message, Ticket state, and Run state are one transaction.
Open a Run detail URL. Show ordered steps, durations, safe tool arguments, citations, and error codes. Explain recursive secret redaction and that replay creates a new linked Run without changing history.
Open Agent Evaluation and run the Mock suite. State clearly that 24/24 is a deterministic contract check, not a model-quality benchmark. Highlight stored dataset hash, commit SHA, mode, timestamp, numerator/denominator metrics, and per-case failures.
Why Go and Eino?
Go makes concurrency, typed boundaries, and small deployable services straightforward. Eino supplies a composable Agent runtime while the project keeps domain rules and observability outside model prompts.
Why a separate MCP server?
It proves protocol-level tool integration and keeps the business boundary independently testable. The tools are intentionally read-only.
How do you reduce hallucination?
Hybrid retrieval, typed business facts, citation requirements, output schema validation, confidence/risk gates, and human review for P0/P1 or weak evidence.
What would change in production?
Replace demo headers with OIDC, add tenant isolation and rate limiting, use a managed vector store/observability backend as scale requires, and establish real-model evaluation thresholds before automated sending.