Create a subagent

A subagent is a bounded specialist with its own instructions, its own short list of data tools, and its own model. This page walks through both ways to create one, the AI-assisted Agent Builder and the field-by-field New subagent form, then shows how to write each field so the main assistant delegates to it well. Creating custom subagents is a Pro-and-above capability. For what subagents are and how a run reads, see Subagents.

Before you begin

Sign in to your account and open the AI Console. You build a subagent from the subagents manager inside it.
Creating custom subagents is available on the Pro plan and above. On Free and Pro Lite the manager still opens, but the create actions are locked and a short line explains that custom subagents are not available on that plan.
The number of custom subagents you can save is capped per tier. When you reach your limit, the create buttons are disabled until you delete one to free a slot. The live tier table shows the current limits.
Optional: to bind your own documents to a subagent, save them to your knowledge first. See Your knowledge and retrieval.

You can also skip the builder and describe the subagent you want directly in chat. The main assistant drafts it from your description, shows you a preview of every setting, and saves it only after you confirm. The two builder paths below remain the full-control routes when you want to set each field yourself.

Create with Agent Builder

Agent Builder drafts a complete subagent from a plain-language description, then hands you every field to review and edit. Use it when you know the job but do not want to hand-pick a model, tools, and a reasoning depth.

Open Agent Builder
In the subagents manager, on the My subagents tab, select Agent Builder.
Describe the job
In What should this agent do?, describe the task in plain language, up to 1,500 characters. Be specific about scope: "Watch my dividend portfolio for cuts and flag them on Mondays" works better than "an agent that helps with stocks".
Add preferences
Optional: in Preferences, add up to 600 characters of steering, for example a request for a low-cost model or for a specific set of tools.
Generate the draft
Select generate. The Builder writes the slug, display name, description, system prompt, model, reasoning depth, allowed tools, max iterations, and web search. Each generation costs a small amount of credits, shown in the product as roughly 0.03 to 0.10 credits.
Review and edit every field
The editor opens with every field pre-filled and memory bindings empty. Review each one and change anything: trim the prompt, swap the model, add or remove tools, adjust a switch.
Save it
Select Create to save. Nothing is stored until you do.

Create with New subagent

The New subagent form starts empty and gives you control of every field from the first keystroke. Use it when you want precise control, or you are cloning a pattern you already know.

Open the empty form
On the My subagents tab, select New subagent. The editor opens with defaults in place: a small, lower-cost model, medium reasoning depth, no tools selected, and direct invocation on.
Name it
Enter a Slug (lowercase letters, digits, and hyphens) and a Display name. The slug is how the main assistant names and picks the subagent.
Write the description
Open the Description card and write, in up to 2,000 characters, when the main assistant should delegate to this subagent and when it should not.
Write the system prompt
Open the System prompt card and write the full instructions the subagent runs under. It never receives your chat history, so include everything it needs.
Choose a model and depth
Pick a model from the picker and set the Reasoning depth. Smaller models cost fewer credits and handle most subagent tasks well.
Grant tools
Open Allowed tools and select the fewest tools the task needs. With none selected, the subagent can answer only from its system prompt.
Set the guardrails
Set Max iterations and, optionally, Max wall seconds so a run cannot spiral. Adjust Memory access, Response size, Direct invocation, and Live web search as needed.
Save it
Select Create. A slug, a display name, a description, and a system prompt are required. If one is missing, the save is blocked with a short message naming it.

Both paths land in the same editor, so the field guidance below applies whichever way you start.

Write each field well

Two facts drive every choice in the form. The main assistant decides whether to delegate by reading your Description, and it does not see the system prompt or the tools when it decides. At run time the subagent receives its System prompt plus the delegated task, and never your chat history. Write the fields with both facts in mind.

Description: This is what the main assistant reads to decide when to hand off, so write it as a routing rule, not a biography. State the triggers ("delegate when the user wants ...") and the exclusions ("do not delegate for data we already have"), and include the words a reader is likely to use. Keep it within the 2,000-character limit.
System prompt: The subagent starts fresh every run, so make the prompt self-contained. Include its role, the inputs it will receive, the exact steps and the order in which to use its tools, and the output shape you want (length, sections, whether to cite sources).
Provider and Model: Pick a model from the picker. When more than one model source is available, a Provider selector sits above the model list and filters it. Smaller models cost fewer credits and answer most subagent tasks well, so reserve heavier models for genuinely hard reasoning. Leaving the model on the default keeps a subagent lean.
Reasoning depth, Max iterations, and Max wall seconds: These are your cost and safety guardrails. A lower depth and fewer iterations produce cheaper, tighter runs, and the wall-time cap stops a single run from spiraling.
Allowed tools: Grant the fewest tools the task needs. Fewer tools mean cheaper runs and less chance of wandering. Bind the subagent to a data area only when the job calls for it.
Memory access: Bind specific saved memories as pinned (always included) or searchable (pulled in when relevant). Leave it empty to let the subagent read all of your memories. See Memory.
Response size: Set Summary when a specialist tends to over-explain and you want a tight readout. Keep Full when the task needs a complete answer.
Direct invocation: Turn this on for a specialist you will want to call by hand with @. Turn it off to let only the main assistant delegate to it.
Live web search: Turn this on only for a subagent whose job is current, outside events. It works on certain newer models only, and when the model you picked cannot use it, the switch locks off and the form warns you.

Test your subagent

After you save, confirm the subagent behaves before you rely on it.

Enable it
Set Enabled on. A disabled subagent stays saved but the main assistant does not see it in the catalog.
Call it directly
In the chat composer, type @ to open the picker and choose your subagent, then send a representative task. Direct invocation must be on for this.
Check the run
Open the credits breakdown under the answer to see the subagent's cost and run time, and open run history to confirm it completed.
Iterate
To refine it, open the subagent and use Suggest updates: describe what should change, then apply or skip each proposed field before you save. Each round costs a small amount of credits, shown as roughly 0.03 to 0.10 credits.

Outputs and how to read them

When the main assistant delegates to your subagent, the subagent's result feeds into the main answer rather than appearing as a separate message. The credits line under a finished answer lists each subagent that ran, with its own cost and run time. Run history shows recent runs and their status, so you can confirm whether a subagent completed, is still running, or failed. For what drives that cost, see GNG credits.

GNG Research provides equity research and educational tools, not investment advice. Nothing on the platform is a recommendation to buy or sell any security. Do your own research and consider your circumstances before making any investment decision.