Custom
Build a custom AI agent with your own instructions, tools and output schema
Custom Agent is a blank frame for jobs the ready-made agents don't cover: you write the main instructions, choose which registry tools it may use and, if you like, give a JSON schema the result must follow.
- As a workflow node
agent.custom- Default model
- Claude Sonnet 5
- Economy model
- Claude Haiku 4.5
- Tools
- 2
- Report
- General
What the agent does
Building a custom AI agent takes no code. Like every other agent, it automatically receives your project's brand voice, the strategy document and, as a saved agent, memory from previous runs; step limits and the run budget apply in the same way. Three ground rules are added to your instructions: gather facts with tools first, never invent data, and only cause side effects (sending, publishing, updating records) when the task explicitly asks for them.
If you pick no tools, the agent works with web search and page reading. All content, comment, knowledge base, messaging, lead, analytics, strategy and platform tools are available to choose from, and the agent can only call the tools on its list. With an output schema, the result comes back as structured data with the same fields every time; without one, you get text.
Typical uses: a product description writer, an incoming message classifier, a competitor price tracker, a brand compliance checker, a weekly industry news digest. For simple jobs a single AI call can handle, the Generate text, Extract structured data and Classify nodes in the workflow editor are cheaper and faster.
Inputs / Outputs
Inputs
- Main instructionsrequired
- Write who the agent is, which rules it follows and what the result looks like, in that order. Example: “You write product descriptions for our cosmetics brand. For each product, write a 70-character title, an 80–120 word description and 3–5 benefit bullets; make no medical claims.”
- Tools it may use
- The capabilities the agent may use: web search, page reading, knowledge base, messaging, lead actions and more. With none selected, web search and page reading are enabled; give publishing, messaging and strategy-update tools only when needed, and with an approval step.
- Output schema (optional)
- A JSON Schema to get the same fields every time, for example
{"type":"object","required":["title","summary"],"properties":{"title":{"type":"string"},"summary":{"type":"string"}}}. Leave it empty to get text.
Outputs
- result
- The agent's deliverable. Without a schema it arrives in this field as text (or a JSON string when structure is needed); with a schema, the output consists of your schema's fields and later steps can address them directly.
Saves a report to the Reports page on every run: General.
Tools
- Web search
web_search - Read a web page
fetch_url
Integrations it uses
Example tasks
- “Write a 70-character title, a 100-word description and 3 benefit bullets for the 5 new moisturizers on our site; make no medical claims.”
- “Pull weekend room prices from the websites of 4 competing hotels in our area; write “none” where you can't find a price and give the result as a table.”
- “Find coffee industry news from the last 7 days, summarize the 5 most important stories in two sentences each and add the source links.”
- “Classify the incoming message as sales, support, complaint or spam, rate its urgency from 1 to 5 and give the reason in one sentence: {{ $json.text }}”
Tips for good results
- Write the main instructions as “who you are → which rules you follow → what the result looks like” and include a sample output.
- Keep the main instructions fixed and put changing details in the task: “Describe these 5 product pages: …”.
- Enable only the tools you need; never give publishing or messaging tools without an approval step.
- With a schema you don't need to describe the format in the instructions. For high-volume uses such as classification, switch off Save report on the node.
- Save it as a saved agent and select it in your workflows; the instructions live in one place and memory works.
Model advice
Default model
Claude Sonnet 5
Per 1M tokens: $2 input · $10 output
Economy model
Claude Haiku 4.5
Per 1M tokens: $1 input · $5 output
The default balanced model is a safe starting point for custom agents that do multi-step research, use tools or write careful copy. For high-volume work such as classification, data extraction or short templated text, the economy model is usually enough and cuts the cost noticeably. For agents that synthesize long documents or support decisions, try a flagship model if your plan allows it.
Frequently asked questions
Do I need to write code to build a custom agent?
No. The main instructions are plain text; the output schema is optional and you can copy and adapt the examples.
How is it different from the ready-made agents?
Ready-made agents come with tuned instructions, tool sets and report formats. With Custom Agent you define these yourself: more flexible, but it takes a few iterations.
Is it safe to give it tools that write or send?
The agent can only call the tools you give it and is told to limit side effects to what the task explicitly asks for. Still pair publishing and messaging tools with an approval step.
Can I write the result to a spreadsheet or CRM?
Yes. With an output schema the result has fixed fields, so you can connect it directly to a node that writes to Google Sheets, HubSpot or Notion.
Get started with OrqLabs
Build your first workflow today
Start on the free plan: connect your accounts, pick a template and stay in control with approval steps. No credit card required.