Subagent

What it does

Spawns autonomous background agents (subagents) that work independently while your main conversation keeps going. Your assistant can fan a single request out into several workers at once: research one thing while it drafts another, investigate a problem off to the side, or pull in a second opinion before committing to an approach.

Each subagent runs in its own context with its own task, so heavy work like deep research, multi-file exploration, and long-running monitoring happens without crowding out your chat. Your assistant decides when delegating is worth it and reports back when a subagent finishes.

Setup required

None. Works immediately.

Permissions

Subagents inherit the same permission rules as your main assistant.
Each subagent gets a role that scopes which tools it can use, so a worker only has the access its task needs.
Only the conversation that started a subagent can check on it, message it, or stop it.

Common prompts

You say...	What happens
“Spawn an agent to research AI startups in healthcare”	Launches a background researcher
“Research the options while you start on the migration”	Runs two subagents in parallel
“Dig into why this build keeps failing”	Delegates to an investigator for root-cause analysis
“Get a second opinion on this plan before we build it”	Consults the advisor and waits for its guidance
“Check on my research agent”	Gets the status of a running subagent
“Cancel the background agent”	Aborts a running subagent

Spawn modes

There are three ways a subagent can start. Your assistant picks the right one for the task, and you can also ask for a specific one.

Mode	How it runs	What it knows
Regular	In the background, in parallel with your conversation	Only the objective and context it's given
Fork	In the background, in parallel with your conversation	Inherits your full conversation (messages, context, and memory)
Advisor	Synchronously (your assistant waits for its answer)	Inherits your full context; runs on a more capable model and returns guidance

Regular.For self-contained work with a clear objective. The subagent doesn't need to know what you've been discussing.
Fork.For work that builds on the conversation so far (“dig deeper on what we just found”). A fork shares your context instead of having it re-explained.
Advisor.A one-shot, read-only second opinion, and the one kind your assistant reaches for on its own judgment: to pressure-test a plan, when it's stuck, or as a final check before calling a task done. It has no tools; it reasons from your context and replies with focused guidance.

Roles

Every subagent runs with a role that determines which tools it can touch. Your assistant picks the most restrictive role that can still do the job, which keeps each worker's blast radius small.

Role	Best for
Researcher	Web and document research, reading and gathering information (read-only)
Coder	Writing and editing files, running commands, build and test work
Planner	Analysis, planning, and synthesizing information (read-only)
Investigator	Root-cause analysis and debugging; returns a compact findings report (read-only)
Advisor	A no-tools, one-shot strategic review
General	Unrestricted access, used only when a task genuinely needs it

Working with a subagent

A subagent follows a simple lifecycle: it starts pending, then runs, and ends completed, failed, or aborted. You don't have to watch for the finish. Your assistant is notified automatically when a subagent ends and follows up with the result.

While one is running, you can stay in the loop:

Check status. Ask how a worker is doing, or what all of them are up to.
Read its output. Pull back what a subagent produced once it finishes.
Send a follow-up. Hand a running subagent new instructions or a course correction.
Cancel it. Stop a subagent you no longer need.
Name it.Give a subagent a memorable label (“the auth research”) so it's easy to refer back to.

Subagents can also reach back on their own. They can surface an interim finding, flag an important result, or signal that they're blocked and need a decision, so you can act on partial progress instead of waiting for the whole task to finish.

Configuration

Silent mode.A subagent's result can be handled internally by your assistant instead of shown to you, for work that's a means to an end. Forks are silent by default.
Model selection. A subagent can run under a specific model profile. By default it inherits the one your conversation is using.
Status tracking. Pending, running, completed, failed, aborted.
Results. Delivered back through your assistant as a follow-up once a subagent finishes.

Tips & gotchas

Genuinely parallel. Regular subagents and forks run in the background while you keep chatting, and you can have several going at once.
Context is a choice.A regular subagent starts fresh and only knows what it's given. Use a fork when the task needs to know what you've been discussing.
Smaller is better. A few focused subagents finish faster and fail more gracefully than one large general-purpose one.
No need to poll.Your assistant is notified automatically when a subagent completes, so you don't have to keep asking.
One level deep.Subagents can't spawn their own subagents, which keeps delegation predictable.
Great for long tasks. Deep research, monitoring, and batch processing are ideal use cases.