Configuration

Configuration Properties

Model Selection Guide

Temperature Reference

Example Configuration JSON

{
  "goal": "Resolve this customer support query: {@ $var.customerMessage }",
  "context": {
    "customer_id": "{@ $var.customerId }",
    "account_tier": "{@ $var.accountTier }",
    "prior_tickets": "{@ $var.priorTickets }"
  },
  "available_tools": [
    "lookupAccount",
    "checkOrderStatus",
    "searchKnowledgeBase",
    "sendEmail"
  ],
  "model": "claude-sonnet-4",
  "max_turns": 8,
  "temperature": 0.2,
  "system_prompt": "You are a BizFirst support agent. Always verify the customer account before performing any action. Respond in a professional, empathetic tone. When you have resolved the query, summarise your actions in JSON format.",
  "credential_id": "anthropic-prod"
}

system_prompt Best Practices

• Use the system_prompt to enforce output format — e.g., "Respond only with valid JSON matching this schema: …"
• Define tool use policies — e.g., "Always look up the account before calling sendEmail"
• Set a persona and tone for customer-facing agents
• Add domain constraints — e.g., "You are processing EU payroll data. Never include PII in tool call metadata"
• Inject fallback instructions — e.g., "If you cannot resolve the query within 5 turns, escalate by calling the escalate tool"

context Field — Detailed Guidance

The context field is separate from the goal and serves as background data the agent receives at the start of each invocation. Use it to provide:

• Structured record data — the customer record, order details, or incident that the goal refers to, so the agent does not need to call a tool just to fetch basic context
• Prior conversation history — if continuing a multi-turn human conversation, pass the prior messages as context
• Environmental constraints — current date, time zone, active promotions, or feature flags that the agent may need to reason about
• Partial data — results from earlier workflow steps that are too large to embed in the goal string

The context is injected as structured data before the goal in the agent's system context. Keep context focused — only include what the agent actually needs. Excessive context increases token consumption across all turns.

Full Configuration Example — Data Enrichment Agent

{
  "goal": "Enrich this lead record with company intelligence and LinkedIn data: {@ $var.leadRecord }",
  "context": {
    "workspace": "{@ $var.workspaceId }",
    "owner": "{@ $var.salesOwner }",
    "crm_source": "Salesforce"
  },
  "available_tools": [
    "lookupLinkedIn",
    "lookupCompanyIntelligence",
    "queryCRM",
    "updateCRMRecord"
  ],
  "model": "claude-sonnet-4",
  "max_turns": 6,
  "temperature": 0.1,
  "system_prompt": "You are a sales intelligence agent. Gather available data on the lead's company and LinkedIn profile. If a source returns no data, mark that field as 'not_found'. Your final output must be valid JSON with fields: company_size (string), industry (string), linkedin_title (string), linkedin_company (string), crm_interactions (integer), enrichment_confidence (0-100 integer).",
  "credential_id": "anthropic-prod"
}

Combining goal and context — Best Practices