Subagents and subagent specialists

Vector's agent does not have to do every part of a task itself. It can hand self-contained pieces of the work to other agents, which run at the same time. There are two kinds, and Vector uses the same two names in the chat, in Background tasks, and in Settings → Agents.

How they differ

What they take on

A subagent takes any self-contained piece of the task. A subagent specialist takes only the kind of work it is named for.

When Vector uses them

Vector's agent decides on its own, and it delegates only when splitting the work costs less than doing it. An independent part that is substantial on its own, roughly five or more files or a long check-and-repair loop, or broad research that would flood its context, goes to a subagent, and several such parts are launched together so they run at the same time while the agent keeps the integration and the final checks. Each subagent starts with a fresh context and re-reads what it needs, so ordinary multi-file changes, fixes, lookups and answers it does itself. It calls a specialist when part of the task fits that specialist's focus. You do not have to ask for either.

How many

A small task gets no subagents; a big one gets one per independent part. Vector sets no cap on either kind. Your computer and your provider's rate limits are the practical limits, and each one spends its own tokens.

While they run

By default the main agent waits for their results, then carries on with what they found. In the desktop app it can also leave one running in the background while it does other work, and picks up the result when that one finishes.

Who they answer to

The main agent. Each one works in its own context and reports back to it once, when it finishes. They cannot ask you questions, though they still ask for permission the way the main agent does, and the main agent tells you what they found.

What they can change

A subagent can read, edit and run commands like the main agent. Explore, Review, Judge and Security can read your repository and run checks, but they cannot edit it. Debug, Test, Performance and Migration can edit.

Where they work

In your session's checkout, not a separate worktree. Vector refuses to start one whose assigned paths overlap another running subagent's.

How deep

One level. By default, subagents cannot start subagents of their own.

In Plan mode

Vector does not start subagents, and of the built-in specialists it uses only Explore, Review and Security, which cannot edit.

Turning them off

Subagents are on by default, and you can turn them off in Settings → Agents or in Vector's config file. Subagent specialists are not affected and keep working.

The eight subagent specialists

Explore, Review, Judge and Security can read your repository and run checks, but they cannot edit it. Debug, Test, Performance and Migration can edit. To ask for one yourself, type @ and its name in the composer.

Where you see them

• In the chatEach batch of delegated work appears as a card in the conversation, with its title, kind, how many agents are working, their status and the time so far, and one square per agent. Click it to open that work in Background tasks.
• In Background tasksOpen the panel with the Tasks button in the session header, or with Cmd+Shift+B (Ctrl+Shift+B on Windows and Linux). Each card shows the kind, the elapsed time, how many agents and their tokens, and one row per agent with its model, tokens and time. A running card has a Stop button, running work is shown in purple, and finished tasks collect under Finished.
• In SettingsSettings → Agents explains both kinds and has the General subagents switch, which turns subagents off. Subagent specialists keep working when it is off.

Turning subagents off

Subagents are on by default. With them off, Vector's agent does the work itself, except parts that fit a subagent specialist, which it can still hand to one. You can also still ask for a specialist yourself with @ and its name.

• In the desktop appSwitch off General subagents in Settings → Agents. Changing it stops any work in progress, and the new setting applies from your next message.
• In the terminalThe terminal reads its own config file, so the desktop switch does not change it. From 1.99.99, add "agent": { "general": { "disable": true } } to ~/.config/vector/vector.json (%USERPROFILE%\.config\vector\vector.json on Windows). It takes effect the next time you start Vector in the terminal.
• For one repositoryFrom 1.99.99, put the same setting in the repository's vector.json. It turns subagents off for that repository in the terminal and the desktop app alike, whatever the switch says. Releases before 1.99.99, such as desktop 1.99.8 and npm CLI 1.99.7, do not read vector.json.

Custom agents

Explore, review, judge and security carry a read-only boundary that is re-applied after your repository's config is merged, so a checked-in file cannot turn a read-only reviewer into one that writes; they can still run verification commands from an allowlist such as git diff, git log and the repository's test and build scripts.

A parent can declare which paths each subagent owns, and Vector refuses two siblings that claim overlapping paths; tasks can depend on each other, run in the background and be resumed later by id.

You can add your own:

vector agent create --description "reviews SQL migrations" --mode subagent vector agent listCopy

Past six running siblings Vector tells the model to mention the spend, because each subagent spends on your keys.

Subagents are not agent workspaces. A subagent works inside your session and your checkout and reports to the main agent. An agent workspace is a separate session in its own worktree or copy that you start, watch in the Agent Dashboard, and merge yourself. See Parallel agents.