Examples

Scenario

A customer sends a message via the support portal. A webhook triggers a BizFirst workflow. The FlowAiAgent node receives the message and the customer ID, then autonomously investigates the account, checks related orders, searches the knowledge base for resolution policies, and sends a resolution email — all without any hard-coded branching.

Configuration

{
  "goal": "Resolve this customer support query professionally: {@ $var.customerMessage }",
  "context": {
    "customer_id": "{@ $var.customerId }",
    "account_tier": "{@ $var.accountTier }",
    "channel": "support_portal"
  },
  "available_tools": [
    "lookupAccount",
    "checkOrderStatus",
    "searchKnowledgeBase",
    "sendEmail",
    "createSupportTicket"
  ],
  "model": "claude-sonnet-4",
  "max_turns": 8,
  "temperature": 0.2,
  "system_prompt": "You are a professional customer support agent for BizFirstGO. Always verify the customer account before taking action. If the query cannot be resolved with available tools, create a support ticket and notify the customer. Your final response must be a JSON object with fields: resolution_summary (string), action_taken (string), escalated (boolean).",
  "credential_id": "anthropic-prod"
}

Expected Agent Behaviour

• Turn 1: Calls lookupAccount with the customer ID to verify identity and tier
• Turn 2: Calls checkOrderStatus if the query mentions an order
• Turn 3: Calls searchKnowledgeBase with the query topic to find resolution policy
• Turn 4: Calls sendEmail with a personalised resolution message
• Terminates with a structured JSON result — no fixed branching needed

Downstream Expression

// Check if escalation was required
{@ $node.supportAgent.result.escalated }

// Get the resolution summary for logging
{@ $node.supportAgent.result.resolution_summary }

Scenario

When a new lead is submitted via a form, the agent enriches the record by pulling data from LinkedIn, a company intelligence API, and the internal CRM. It then synthesises all data into a structured enrichment object and updates the CRM record.

Configuration

{
  "goal": "Enrich this sales lead with company intelligence, LinkedIn profile data, and CRM history. Lead details: {@ $var.leadRecord }",
  "context": {
    "workspace_id": "{@ $var.workspaceId }",
    "sales_owner": "{@ $var.salesOwner }"
  },
  "available_tools": [
    "searchLinkedIn",
    "lookupCompanyIntelligence",
    "queryCRM",
    "updateCRMRecord"
  ],
  "model": "claude-sonnet-4",
  "max_turns": 6,
  "temperature": 0.1,
  "system_prompt": "You are a sales intelligence agent. Gather all available data and produce a structured enrichment. If any data source returns no result, note that field as 'not_found'. Your final response must be valid JSON with fields: company_size, industry, linkedin_title, linkedin_company, crm_prior_interactions, enrichment_confidence (0-100).",
  "credential_id": "anthropic-prod"
}

Typical Tool Call Sequence

Scenario

Inbound requests arrive via email or form. The agent classifies the request type, determines the appropriate department and priority, and routes it to the correct queue — replacing a rigid keyword-matching rule set with flexible AI reasoning.

Configuration

{
  "goal": "Classify this incoming request and route it to the correct internal queue: {@ $var.requestText }",
  "context": {
    "sender_email": "{@ $var.senderEmail }",
    "received_at": "{@ $var.timestamp }",
    "channel": "{@ $var.channel }"
  },
  "available_tools": [
    "classifyIntent",
    "lookupSenderProfile",
    "checkQueueCapacity",
    "routeToQueue",
    "sendAcknowledgement"
  ],
  "model": "claude-haiku-3-5",
  "max_turns": 5,
  "temperature": 0.0,
  "system_prompt": "You are a triage agent. Classify requests into one of: billing, technical_support, sales_inquiry, hr_query, compliance. Set priority as: urgent, high, normal, or low based on keywords and sender history. Always send an acknowledgement after routing. Output JSON: {\"category\": \"...\", \"priority\": \"...\", \"queue_id\": \"...\", \"acknowledgement_sent\": true/false}.",
  "credential_id": "anthropic-prod"
}

Scenario

Every Monday morning a scheduled trigger fires. The FlowAiAgent node collects KPI data from multiple internal APIs, performs trend analysis, and produces a formatted markdown report that is then stored to S3 and emailed to the management team.

Configuration

{
  "goal": "Generate the weekly management summary report for the week ending {@ $var.weekEndDate }. Collect all required KPIs, identify trends, and write a concise executive summary followed by department sections.",
  "context": {
    "report_recipients": "{@ $var.recipients }",
    "regions": ["APAC", "EMEA", "Americas"],
    "currency": "USD"
  },
  "available_tools": [
    "querySalesDB",
    "queryFinanceAPI",
    "queryOperationsAPI",
    "calculateTrend",
    "writeReportToS3",
    "sendEmailWithAttachment"
  ],
  "model": "claude-opus-4",
  "max_turns": 15,
  "temperature": 0.4,
  "system_prompt": "You are an executive business analyst. Produce a professional weekly report in markdown format. Structure: 1) Executive Summary (3-5 bullet points), 2) Sales Performance by region, 3) Operations KPIs, 4) Financial Summary, 5) Key Actions Required. Use tables where appropriate. After writing the report, save it and send it.",
  "credential_id": "anthropic-prod"
}

Why claude-opus-4 Here?

Report generation requires synthesising large amounts of numeric data, identifying multi-period trends, and producing well-structured prose in a single coherent document. Opus-4's extended reasoning and writing capability produces significantly higher-quality reports than Sonnet or Haiku for this task, and the higher token cost is justified by the strategic value of the output.

Downstream Workflow Steps

• After success port: send Slack notification with S3 link to report
• After error port: alert the operations team that the report failed and log the partial_reasoning for diagnosis

// Get the S3 URL of the stored report from the tool_calls output
{@ $node.reportAgent.tool_calls | filter(x => x.tool_name == "writeReportToS3") | first | .output.s3_url }