Customer Operations Agent

Grounded customer operations agent with an escalation skill, deterministic tools, and approved runbook knowledge.

Project overview

Veryfront Code organizes these primitives with clear source-tree conventions for agents, skills, knowledge, tools, app routes, evals, workflows, schedules, and webhooks.

customer-operations-agent/
  agents/        # agent definitions
  skills/        # reusable agent instructions
  knowledge/     # approved runbooks
  tools/         # deterministic tools
  app/           # chat UI and AG-UI route
  evals/         # retrieval and grounding checks
  workflows/     # repeatable escalation operations
  schedules/     # source-defined scheduled runs
  webhooks/      # source-defined event runs

Prerequisites

• Node.js and npm.

Getting started

Install dependencies:

npm install

Sign in to the Veryfront Cloud gateway so chat, evals, workflows, schedules, and webhooks can access models:

npx veryfront login

Start the chat UI:

npm run dev -- --port 3010

Open http://localhost:3010.

Try the agent

Ask support questions that match the included runbooks:

• Users cannot sign in with SSO after yesterday's deployment. Production support is blocked.
• The customer's renewal invoice failed payment, but the workspace is still active.
• A migration shipped this morning and users now see errors in the onboarding workflow.
• One support manager cannot access the correct workspace after changing browsers.

Validate the agent

Run the eval suite after signing in.

npm run eval

The eval suite checks tool use, retrieval quality, and grounded answers:

• agent.calledTool("search_knowledge")
• agent.noFailedTools()
• knowledge.recallAtK
• knowledge.precisionAtK
• knowledge.mrr
• answer.groundedness

Reports are written to timestamped folders under .veryfront/evals/. Each dataset row declares metadata.expectedKnowledge, so retrieval quality is measured against the runbooks the case should use.

Compare models with explicit baseline and candidate models.

npx veryfront eval support-triage \
  --baseline-model anthropic/claude-sonnet-4-6 \
  --candidate-model moonshotai/kimi-k2.6 \
  --candidate-model openai/gpt-5.4-nano \
  --json

The comparison report writes per-model results plus comparison.json and comparison.md in the timestamped report folder.

Use custom provider credentials only with matching model settings.

Automate agent

Workflow

Test the escalate-ticket workflow with fixture input.

npm run workflow:run

This checks that support-agent can draft scope, evidence, owner, and next action.

Schedule

Discover the source-defined schedule, then test it locally.

npm run schedules
npm run schedule:run

Webhook

Discover the source-defined webhook, then test it locally.

npm run webhooks
npm run webhook:run

The schedule and webhook both target the same escalate-ticket workflow. In Veryfront Cloud, deployment creates or updates the hosted schedule and webhook from these source files.

Deployment

Sync local files to the main branch of your Veryfront project without deploying.

npx veryfront push --branch main

Deploy the main branch to the preview environment.

npx veryfront deploy --branch main --env preview