Build a Search Assistant

Create the workflow and wire Input to Output

Create a new workflow. Drag an Agent node onto the canvas between the Input and Output nodes, and connect Input → Agent → Output. On the Input node, add a single input field named question — that is what callers send.

Pick the agent's model

Select the Agent node and choose a model under LLM. Any LLM provider available in your project works — use your default LLM connection; the gear icon next to the selector opens the model's own settings if you want to adjust temperature or the connection. See The Agent Node for the full configuration reference.

Attach a web-search tool

In the Agent panel's Tools section, click Add tool and pick Web search with Tavily from the catalog.

The tool attaches as a child node under the agent. To use a different provider, choose Web search with Exa, Web search with ScaleSerp, or Web search with Firecrawl instead — each has its own node reference and Connection type.

Create the Tavily Connection

Click the gear icon on the attached tool to set its Connection. Create a new Tavily Connection (it needs only an api_key) or pick an existing one — the full flow is in Create a Connection. The tool cannot run until it points at an active Connection.

Give the agent its role

In Role & Instructions, paste a role that tells the agent to refine the query, search, and cite its sources:

You are a web research assistant. For every question:

1. Refine the request into one or two precise search queries before searching.
2. Use the web search tool to gather current information. Search again with
   different terms if the first results are thin or off-topic.
3. Answer only from what the search results support. If the results do not
   answer the question, say so plainly instead of guessing.
4. Write a concise, direct answer, then a "Sources" list with the title and URL
   of every page you relied on. Cite each claim inline with a bracketed number
   that maps to that list.

The role field accepts Jinja, so you can inject workflow inputs into it later — see Prompts, Roles & Inference Modes.

Test it

Open the Test panel, enter a question that needs current information, and run it.

Watch the result: the agent should search, then return a cited answer. Open the trace to see each search query and the results the agent read — if it answered without searching or cited nothing, tighten the role.

Add the orchestrator and its Manager LLM

Drag Graph Agent Orchestrator from the AGENTS section of the palette onto the canvas, between Input and Output. Drop an LLM node onto its Add LLM here placeholder to set the Manager LLM — a fast, inexpensive model is right here, since the manager only makes small routing and input-shaping decisions.

(The screenshot shows the graph-researcher template's research and validate states; yours will be named search and write.)

Create the search and write states

Open the orchestrator's panel and add two states under Nodes, renaming them search and write. Give each an Agent task under its Tasks section.

In search, attach Web search with Tavily to the agent (as in Pattern 1) and give it a researcher role:

You are the researcher on a two-agent team. Given a question, produce the raw
material another agent will write from.

1. Break the question into the specific facts you need to find.
2. Use the web search tool to gather them, searching multiple times with
   different queries until you have enough.
3. Return your findings as bullet points. After each fact, include the source
   title and URL it came from. Do not write prose or a final answer — that is
   the writer's job.

In write, the agent has no tools; it works only from the researcher's notes:

You are the writer on a two-agent team. You receive research notes with sources
and turn them into the final answer. You have no tools — work only from the
notes you are given.

Produce:
- A two to three sentence summary that answers the question directly.
- A short section of supporting detail, with each claim cited inline as a
  bracketed number.
- A "Sources" list mapping each number to its title and URL.

If the notes do not support an answer, say what is missing rather than inventing
detail.

Wire the states and test

In the orchestrator's Edges section, replace the default START → END edge so the flow is START → search, search → write, write → END. Connect Input → orchestrator → Output, map the orchestrator's Input to $.input.output.question, and run it from the Test panel. The orchestrator returns the writer's final message as its content output.

Rephrase the question

As in Pattern 1, the Input node has a single field named question. Add an LLM node after Input, and rename it rephrase (selectors bind by node name, so name it before mapping anything downstream). Give it a prompt that emits only a search query:

Rewrite the user's question as a single, focused web-search query. Return only
the query text, with no quotation marks or commentary.

Question: {{ question }}

The {{ question }} placeholder becomes an input field — map it to $.input.output.question.

Search with the rephrased query

Add a Web search with Tavily node after the rephrase node, rename it websearch, and set its Tavily Connection as in Pattern 1. Map its Query input to the rephraser's output:

$.rephrase.output.content

Write the cited answer

Add a second LLM node — the writer — after the websearch node. Its prompt takes the question and the search results:

Answer the question using only the search results below. Cite each claim inline
with a bracketed number, and end with a "Sources" list mapping each number to
its title and URL. If the results do not answer the question, say so.

Question: {{ question }}

Search results:
{{ results }}

Map the two placeholders:

Connect the writer to Output, then test. Because there is no agent loop, a typo in a selector fails quietly — a mapped field that finds no match resolves to null. If an answer comes back empty, open the trace and check each node's resolved input against How nodes connect and the selector rules.