Cookbook

The rest of these docs document Wetel’s primitives — agents, workflow nodes, knowledge bases, MCP connectors — one at a time. A recipe is different: it’s a complete, working configuration for a specific real-world use case, built entirely out of those same primitives, shown end to end from “create the agent” through “here’s what happens when a user talks to it.”

Recipes are meant to be copied and adapted, not read for background. If you already know which primitive you need (a condition node, knowledgeBaseId, a webhook node), the Workflows section and API Reference are the right place. If you’re starting a new agent for a specific job and want a proven shape to start from, start here instead.

For a lighter, more abstract tour of individual node combinations, see Worked Examples & Demos — the recipes below go deeper on one full use case each, rather than illustrating a single pattern in isolation.

Recipes

• Support Agent with a Knowledge Base (RAG) — ground an agent’s answers in your own documentation: create a knowledge base, ingest a FAQ, and wire it into a workflow so the agent answers from your content instead of guessing.
• First-Round Candidate Screening Agent — an interview agent that asks a fixed number of questions, closes professionally, and produces a structured hire/no-hire evaluation your own systems can pull.
• Order-Status & Product-Support Agent — combine a product-catalog knowledge base with a live order-lookup webhook call, so the same agent can answer “what’s this warranty cover?” and “where’s my order?” in one conversation.
• Read Channel Facts Without an LLM Call — replace the llm nodes that copy ids, names and contact details out of a message with one set_variable node reading channel.raw, with a measured before/after (about 55% fewer tokens in the example run).
• Read Many Photos in One Node — a step-by-step, live-verified build where one llm node reads up to 40 phone photos sent as [image:<url>] markers, with the trace fields, limits and troubleshooting.
• Triage a Support Ticket End to End — one complete, pasteable graph that collects a ticket with await_reply, classifies it, normalises the values with set_variable, branches with a switch, asks a workspace owner to approve urgent escalations, replies and ends the session.
• Answer When Mentioned in Group Chats — put an agent into your platform’s group chats so it answers only when someone mentions it or replies to it: one session per group, safe sender context with turnContext, and your own caps.
• Fail Gracefully — keep the conversation going when a partner API call fails: success/failure edges, a switch on _lastNodeErrorCode, and only _lastNodeErrorPublic shown to the user.
• Build a Reliable Workflow — retry only what is safe to resend, route what is left to failure edges with safe error text, set a workflow-level failureResponse, prove every branch in mock mode, and publish a version you can roll back. Available on production from 2026-10-23.
• Delegate to a Specialist — hand one question to a specialist agent with sub_agent, pass named values with inputs, get structured values back with outputs, and branch on them.
• Delegate Work to a Colleague and Keep the Owner in Control — a manager employee plans, asks the owner to confirm, delegates to a specialist that drafts and stops at an owner approval; request changes up to 3 times, approve to save a page to Documents. Engine support in production since 2026-11-05; the example employees and screens are a demo on staging.
• Provision Agents for Your Customers With a Template Hire — hire a ready-made role per customer with a templates:hire key, idempotently, bind shared resources, and start sessions on it.

More recipes will be added over time as new use-case patterns prove out.