AI Agents — Examples and recipes

Ready-to-build recipes. Each one states which nodes to place, how to connect them and the key fields. For the detail of each field, see the node reference.

---

1. Basic agent (brain only)

An assistant that answers questions, with no tools or external knowledge. The “hello world” of agents.

Nodes: Start → AI Agent → End, with a Brain (LLM) connected to the brain port.

1. AI Agent — system_prompt: “You are a Floogos support assistant. Answer briefly.” input: leave it empty to use whatever arrives from the flow.

2. Brain (LLM) — provider: openai, model: gpt-4o, credentials_id: your OpenAI credential. Connect its output to the brain port.

Result: the agent answers in data.response. It’s the base you add the other capabilities on top of.

---

2. Agent with one tool (generate images)

The agent can generate an image when the user asks, using the openaiImages node as a tool.

Nodes: the basic one + a Tool connected to the tools port.

1. Tool — configure:

   • tool_name: generate_image
   • tool_description: “Generates an image from a text description. Use it when the user asks to create, draw or illustrate something.”
   • tool_kind: node
   • node_type: openaiImages
   • tool_schema:

     {
       "type": "object",
       "properties": {
         "prompt": { "type": "string", "description": "Description of the image to generate" }
       },
       "required": ["prompt"]
     }

2. Connect its output to the agent’s tools port.

What happens: if the user writes “draw me an astronaut cat”, the LLM decides to call generate_image with { prompt: "astronaut cat" }, the agent runs openaiImages and returns the result to the LLM so it can answer.

Tip

The same pattern works for any node: http (call an API), finnegansGet (read an ERP), googleSheets (read/write a spreadsheet), etc. You change node_type and the tool_schema.

---

3. Agent with RAG (answer about your documents)

An assistant that answers based on your documentation. It’s two workflows: one that indexes (once, or when the docs change) and another with the agent that answers.

3a. Ingestion workflow (index)

Nodes: Start → (node that brings the text, e.g. reads a file or a table) → RAG (ingest) → End.

1. RAG (ingest) — namespace: product-manual, provider: openai, model: text-embedding-3-small, credentials_id: your credential, chunk_size: 1000, overlap: 150.

2. The input (data.text or data.documents[]) is the content to index. Run the workflow once.

3b. Agent workflow (retrieve)

Nodes: the basic agent + a RAG (retrieve) connected to the rag port.

1. RAG (retrieve) — namespace: product-manual (the same as the ingestion), top_k: 4, provider: openai, model: text-embedding-3-small (the same model), credentials_id: your credential. Connect its output to the rag port.

2. AI Agent — system_prompt: “Answer using only the retrieved context. If it isn’t there, say so.”

What happens: for each question, the agent retrieves the 4 most similar chunks from the namespace and injects them into the prompt before answering. That way it answers with your content and doesn’t make things up.

Caution

The embeddings namespace and model must be identical in ingestion and retrieval. It’s the most common RAG mistake.

---

4. Agent with skills (narrow behavior)

A support agent that has a billing skill which only activates when the request warrants it, and which enables a specific tool.

Nodes: the basic one + a Tool (issue_invoice) + a Skill connected to the skills port.

1. Tool issue_invoice — point it at your billing node/workflow. Connect it to the tools port.

2. Skill — configure:

   • skill_name: billing
   • description: “Issuing invoices and receipts, VAT queries and tax data.”
   • keywords: invoice, receipt, vat, tax
   • prompt_fragment: “When issuing an invoice, first confirm the customer’s tax details.”
   • enables_tools: issue_invoice
   • Connect it to the skills port.

What happens:

• If the user asks something general, the skill does not activate and the issue_invoice tool is not offered to the LLM (it stays gated).
• If the user mentions “invoice” or “VAT”, the skill activates: it injects its instruction and enables issue_invoice.

Note

This gating keeps the prompt light and prevents the LLM from using tools out of context. A tool referenced by a skill only appears when its skill activates; tools without a skill are always available.

---

5. Agent with a local LLM (Ollama / vLLM)

An agent that runs against a local model, without sending data to an external provider — ideal for on-premise.

1. Create a local credential with your server’s base URL (e.g. http://localhost:11434/v1 for Ollama).

2. Brain (LLM) — provider: local, model: whatever you have loaded (e.g. llama3.1), credentials_id: the local credential. Connect to the brain port.

What happens: the agent talks to your local server using the OpenAI-compatible API. Everything else (tools, RAG, skills) works the same.

Tip

For RAG with local embeddings, in the RAG nodes choose provider: local and the embeddings model your Ollama/vLLM serves.

---

Next step

Something not answering as expected? See common problems and solutions.