Building Workflows

Create multi-step integrations across your tools

This guide covers how to build workflows that span multiple integrations – from simple queries to multi-step pipelines with type bridging and approval gates.

Simple Query

The fastest way to call a tool:

Playground

Type a goal and the system finds and executes the right tool:

You: Show me my Salesforce leads
System: [discovers salesforce@1/get_leads@1, executes, shows results]

Conduit SDK

result = client.perform("salesforce@1/get_leads@1", {"limit": 10})

Sandbox

Open the integration’s detail page, go to the Sandbox tab, select the tool, fill in parameters, and click Run.

Multi-Step Workflow

When your goal involves multiple systems, the planning engine chains tools automatically.

Via Playground

You: Create QuickBooks invoices for qualified Salesforce leads

The system builds a plan:

  1. salesforce@1/get_leads@1 (filter: status = Qualified)
  2. data-grout@1/prism.focus@1 (convert lead to customer)
  3. quickbooks@1/create_invoice@1

Review the plan and approve to execute. Each step runs in sequence, with type bridging handled automatically.

Via the API

Use discovery.plan to generate a verified plan, then flow.into to execute it:

plan = client.call_tool("data-grout@1/discovery.plan@1", {
    "goal": "Create invoices from qualified leads"
})

result = client.call_tool("data-grout@1/flow.into@1", {
    "plan": plan["steps"],
    "save_as_skill": True,
    "name": "lead_to_invoice"
})

The plan comes with a Cognitive Trust Certificate proving it is cycle-free, type-safe, and policy-compliant before any tool runs.

Type Bridging

When tools produce data in one format and the next tool expects another, the planner inserts prism.focus to convert between types:

salesforce@1/get_leads@1 → crm.lead@1
  ↓ prism.focus (crm.lead@1 → billing.customer@1)
quickbooks@1/create_invoice@1 ← billing.customer@1

This happens automatically during planning. You can also call prism.focus directly:

customer = client.call_tool("data-grout@1/prism.focus@1", {
    "data": lead,
    "source_type": "crm.lead@1",
    "target_type": "billing.customer@1"
})

Approval Gates

For operations that modify data, insert an approval gate:

approval = client.call_tool("data-grout@1/flow.request-approval@1", {
    "action": "Create 25 invoices in QuickBooks",
    "details": {"total_amount": 47500, "count": 25},
    "reason": "Bulk invoice creation"
})

if approval["decision"] == "approved":
    # proceed with execution

Approval gates pause the workflow until a human responds. Configure which operations require approval in your Policy settings.

Saving Skills

After a successful workflow, save it as a reusable skill:

result = client.call_tool("data-grout@1/flow.into@1", {
    "plan": [...],
    "save_as_skill": True,
    "name": "lead_to_invoice"
})

Skills retain their CTC and execute without re-verification on subsequent runs:

result = client.perform("skills@1/lead_to_invoice@1", {})

Error Handling

When a tool call fails, the response includes a structured error:

{
  "error": {
    "code": "upstream_error",
    "message": "QuickBooks API rate limit exceeded"
  }
}

Common error codes: tool_not_found, invalid_arguments, upstream_error, timeout, rate_limit, policy_violation.

Check the Runs tab for detailed execution logs, or use the Agent Inspector for cognitive-level debugging.

Related