Subagents

Subagents are specialist helpers you create or import. The main assistant can delegate a task to one, or you can call one directly, so a focused job runs with its own instructions and its own set of data tools. This page explains what the subagent area shows, the fields in the builder form, how to read a run, and what drives its credit cost. Subagents live inside the AI Console. The area opens on every plan, but creating your own subagents and using the built-in specialists both require the Pro plan or above. On the Free and Pro Lite plans the area stays visible with the create actions locked, and the number of custom subagents you can save is capped per tier above the gate. For a step-by-step walkthrough of both creation paths, see Create a subagent.

Before you begin

Sign in to your account. Subagents run inside the AI Console.
To bind your own documents to a subagent, save them to your knowledge first. See Your knowledge and retrieval.

Quickstart

Open the subagents manager
Go to the AI Console and open Agents from the console sidebar. This entry is the subagents manager.
Create a subagent
Open the builder form and set a slug, a display name, a description, and a system prompt.
Choose a model and tools
Pick a model and reasoning depth, then select the data tools the subagent may use.
Save and enable it
Save the subagent and set Enabled on. The main assistant can now delegate to it, and you can call it with @ if direct invocation is on.

What it shows

The subagents manager opens from the console sidebar entry labeled Agents, with three tabs plus a per-subagent editor. The parts are described in the following sections.

My subagents

The specialists you have created, plus Agent Builder (describe a task and let the platform draft the subagent for you) and New subagent (build one field by field). On the Pro plan and above, a counter shows how many of your tier's subagent slots you have used. On the Free and Pro Lite plans the create buttons are locked, and a short line explains that custom subagents are not available on that plan.

Built-in

Ready-made specialists the platform provides, available on the Pro plan and above, each shown with its description and the data tools it can use. More appear here as the platform enables them. One built-in is available today: Web Researcher, which researches current news and events on the live web and returns a synthesis with inline citations.

Marketplace

Browse, import, and publish community subagents. This is a separate surface, covered in Marketplace.

The editor and the explainer

A per-subagent editor holds all of its settings, an AI Suggest updates helper for an existing subagent, and a Publish action to share it. A short in-product explainer states that subagents are bounded specialists the main agent delegates to by reading the description you write. A subagent cannot draw inline widgets and cannot delegate further, and every subagent run is metered against your GNG Credits.

How to invoke a subagent

You reach a subagent in two ways, described in the following list.

Delegation: the main agent reads each subagent's description and decides on its own when to hand off a subtask. Write the description around the decision, for example "delegate when" and "do not delegate if". The main assistant can hand several sub-tasks to different subagents at once; those runs execute together, and the credits line lists each one separately with its own cost and run time.
Direct mention: in the composer, type @ (or select the @ button) to open the picker, then choose a subagent. The chat shows a Routing to row naming the agent, and your next message goes straight to it. Select Main agent on that row to switch back. Direct mention works only for subagents that have direct invocation turned on.

Manage agents from chat

You can also ask the main assistant to create or update a subagent in plain English, without opening the builder. Describe the change you want, and the assistant drafts it, shows you a preview of the exact settings, and waits for your confirmation in chat before it saves. Low-risk tweaks, such as switching an agent on or off, apply straight away. Deleting a subagent stays in the manager and is deliberately not available in chat.

Inputs

The builder form sets how a subagent behaves. The fields are described in the following list.

Slug: lowercase letters, digits, and hyphens. This is the short name the main assistant uses to pick the subagent.
Display name: a readable name.
Description (up to 2,000 characters): tells the main assistant when to delegate to this subagent and when not to. The main agent reads this to decide on a hand-off, so write it around the decision.
System prompt: the full instructions the subagent sees. It does not receive your chat history, so include everything it needs.
Provider: shown above the model when more than one model provider is available. It sets which family of models the model list offers; switching provider reselects a recent eligible model from that provider.
Model: chosen from the available list. Defaults to a small, lower-cost model. Smaller variants usually save credits and answer most subagent tasks well.
Reasoning depth: Low (cheapest), Medium, High, or Max. Default Medium.
Max iterations: 1 to 20, default 20. The number of tool-use steps the subagent may take. More steps mean more cost.
Max wall seconds: optional. A hard time cap for one run. Blank uses the platform default of thirty minutes.
Allowed tools: the data tools the subagent may use, grouped by topic. With none selected, the subagent answers only from its system prompt.
Skills: attach your saved skills to a subagent so it always carries those instructions, available when skills are on for your account. See Skills.
Memory access (your subagents only): bind specific memory files to put the subagent on a curated list, as pinned (always included) or searchable (pulled in when relevant). Leave it empty to let the subagent read all of your memories. See Memory.
Enabled: a saved-but-hidden switch. When off, the subagent stays saved but the main assistant does not see it.
Response size: Full (model default) or Summary (a tighter reply of about 1,500 tokens). Summary keeps replies tight.
Direct invocation enabled: when on, you can @ the subagent to chat with it one to one; when off, only the main assistant can delegate to it.
Live web search: lets the subagent search the web and cite its sources during a run. It applies only to certain newer models; the form warns you when the chosen model cannot use it.

Outputs and how to read them

When you use a saved subagent in AI Workflows, its available tools can differ from chat. Check X of Y tools run in workflows in the picker. Supply missing information through connected data steps and select the needed fields before testing.

When a subagent runs, its result feeds into the main answer. The credits line lists each subagent that ran with its own cost and run time. A subagent keeps working until its task is done or it reaches its own limits (its time cap and step count), rather than stopping mid-answer. Run history shows recent runs and their status, so you can confirm whether a subagent completed, is still running, or failed.

How it is calculated

There is no scoring formula behind a subagent. Each run is metered on its own and reported separately in the answer's credits breakdown, so you always see what a given subagent cost. The settings that raise that cost are a higher reasoning depth, more iterations, a heavier model, and web search. To keep cost down, prefer a smaller model, a lower depth, and the fewest iterations and tools the task needs. For the user-facing meaning of the meter, see GNG credits.

Data and timing

A subagent uses whichever data tools you allow, on the same live data as the main chat, sourced from AlphaVantage (Nasdaq quotes delivered through AlphaVantage) and from EDGAR, plus any of your own knowledge you bind to it. Data is current to the platform's normal refresh schedule.

Use cases

Build a market-brief helper that pulls recent price action and news for a set of tickers.
Create a focused fundamentals reviewer that reads only the financial statements you care about.
Bind a subagent to a specific uploaded document so it answers from that source.
Keep a subagent disabled while you refine it, then enable it when ready.

Limitations and disclosures

Subagent output is research support, not a recommendation. Creating your own subagents and using the built-in specialists both require the Pro plan or above; on the Free and Pro Lite plans the subagent area is visible but the create actions are locked. Above the gate, the number of custom subagents you can save is capped per tier, and the builder shows your current count against your limit. Running a subagent draws down 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.