My first agent
From blank canvas to published agent: four cards, three connections and an assistant that searches the web.
Updated on Aug 10, 2026
Let's build the most useful agent possible with the minimum pieces: a web search assistant that researches before answering. Four cards and three connections — here's the result on the canvas:
- 1
Create the agent
On platform.hinow.ai, Create → Agents → New agent. The canvas opens with the Start card already there — it's fixed, every flow begins with it.
- 2
Drag the Agent
From the palette on the left, drag the Agent card below Start. Click it to open the modal: name it
Researcher, choose the model (we usehinow/himax, our in-house model) and write the prompt — see below. - 3
Drag Web Search
Drag Web Search to the right side of the Agent. In the modal, leave
num_results: 5(the default works fine). - 4
Make the three connections
Output of Start → top of Agent (flow). Left handle of Web Search → slot on the right of Agent (tool, the dashed line). Output of Agent → top of End.
- 5
Test and publish
Use the editor's test chat ("what's the dollar rate today?") and watch the agent decide to call the search. Did it work? Publish — the published version is what the API serves.
Only two cards need real configuration in this agent:
{
"name": "Pesquisador",
"model": "hinow/himax",
"system_prompt": "Você é um assistente de pesquisa. Quando a pergunta envolver fatos atuais (notícias, preços, eventos), USE a ferramenta de busca antes de responder. Cite as fontes que usar. Se a pergunta for atemporal, responda direto.",
"config": {
"temperature": 0.7,
"max_tokens": 4096,
"max_tool_loops": 10
}
}{
"config": {
"num_results": 5,
"safe_search": true
}
}The prompt is the tool manual
The model only calls the search when it understands it should. Stating when to use ("current facts") and what to do with the result ("cite sources") in the prompt matters more than any configuration.
The three lines on the canvas have two different roles — understanding this is understanding the entire builder:
| Connection | Type | What it means |
|---|---|---|
| Start → Researcher | Flow (solid line) | Execution order: the user's message enters and the turn begins at the agent. |
| Web Search → Researcher's slot | Tool (dashed) | NOT executed in sequence — it becomes a *tool* that the model can call as many times as it wants during the turn (up to max_tool_loops). |
| Researcher → End | Flow (solid line) | When the agent finishes the response, the turn ends with status success. |
In execution, a typical turn looks like this: the question arrives through Start → the Researcher reads the prompt and decides to call web_search("dollar rate today") → receives the 5 results → writes the response citing sources → the flow continues to End. In the event stream you see exactly this story: workflow_node_start → tool_use → tool_result → agent_message → done.
Everything you built becomes a workflow {nodes, edges} saved in the agent — this is what the API executes after publish:
{
"nodes": [
{ "id": "start", "type": "start", "data": { "label": "Início", "variables": [] } },
{ "id": "agent-1", "type": "agent", "data": { "name": "Pesquisador",
"model": "hinow/himax", "system_prompt": "Você é um assistente de pesquisa..." } },
{ "id": "search-1", "type": "web_search", "data": { "config": { "num_results": 5 } } },
{ "id": "end-1", "type": "end", "data": { "label": "Fim", "status": "success" } }
],
"edges": [
{ "source": "start", "target": "agent-1" },
{ "source": "search-1", "sourceHandle": "tool", "target": "agent-1", "targetHandle": "slot-1" },
{ "source": "agent-1", "sourceHandle": "output", "target": "end-1" }
]
}Notice the targetHandle: "slot-1" — this is how the runtime knows that Web Search is a tool of the agent, not a step in the flow.
curl -N -X POST https://agents.hinow.ai/v1/agents/SEU_AGENT_ID/run \
-H "Authorization: Bearer hi_SUA_API_KEY" \
-H "Content-Type: application/json" \
-d '{"message": "qual a cotação do dólar hoje?"}'