본문으로 건너뛰기

Work with subagents

A subagent is a separate conversation with its own assignment, tools and recorded result. The main agent remains responsible for integrating its work.

Ask the agent to delegate​

In a supported native Desktop build, ask the agent to split independent work among subagents. For example: “Use five subagents to investigate these five questions, then combine their findings.” Ordinary Act work can delegate at every effort level; choosing Ultra is not required. A headless client opts in with --subagents.

Each worker has a stable identity and appears in the parent's Subagents view. It receives its assignment and captured project guidance and skills, rather than an automatic copy of the parent conversation. It uses the parent's captured model, reasoning setting and connection; changing the parent's selection later does not move accepted child work to another connection.

Workers can use the native file and command capabilities explicitly delegated from the parent's accepted permission profile. They cannot gain wider access, change agent settings, inherit live editor connections, or delegate again. An action needing additional permission is refused in the worker and must be reported to the main agent; it does not create an invisible approval wait.

One parent turn can create up to eight workers, each with up to eight model requests. This reserves a separate bounded allowance of 64 child requests beside the parent's work allowance; it is not a token or monetary spending limit. Waiting observes the assigned work and returns its recorded result. Information sent to an active worker joins its current assignment and becomes available to its next model request. The receipt confirms that the information was recorded, not that the model has read or acted on it. Sending refuses paused or finished work and never restarts an idle child. The main agent must collect the assignments' results before reporting successful completion. General workers belong to the parent turn that creates them and receive one bounded assignment. The model cannot reuse them for follow-up work; its messaging, waiting and stopping tools address only workers from that same parent turn. A new assignment requires a new worker, while earlier child conversations remain available to inspect. The main chat's Stop also stops its owned workers.

Processes a worker starts belong to its assignment. The native host automatically stops any remaining processes before its successful result is collected. A worker can leave a bounded process running when the assignment requires it and report its session and last observed state; work intended to keep running must be started by the main agent. Waiting observes this cleanup rather than causing it. A missing or unverified cleanup record remains explicit uncertainty, including after restart. The assignment's result and resource retirement are reported separately.

General workers are available for ordinary Act turns on connections supporting JSON tools. Plan, persistent Goals, temporary side chats and children retain their separate capability limits. Goals keep their existing funded reviewer without acquiring an additional general-worker allowance.

Open subagent activity​

In a supported Grida Desktop build on macOS or Linux, click the subagent indicator in the parent transcript or a named start/finish event. In a workspace, this opens Subagents in the main content tabs, beside your documents. Requesting it again brings the same tab forward. A standalone file window shows it within the chat instead.

The view belongs to the selected parent chat. Active includes ready, queued, running, paused and uncertain work; Done keeps completed, failed and stopped work when no assignment is still waiting.

The view is read-only. It shows subagent status, activity and conversations; creation, assignment, steering and other execution controls belong to the main agent. Give further instructions in the parent chat.

Select a child to read its conversation and activity. Back returns to the list in the same tab. The parent chat and composer remain visible and usable, with the unsent draft intact. The child's name and visual identity match its links in the parent's activity. Start and finish events belong to the parent turn that sponsored the assignment. On compatible native builds, they appear between the parent's commentary and tool activity where the host recorded them, rather than collecting at the bottom. Placement stays consistent during live updates and after reopening. Reading an event or reopening a child starts no model request.

Older history or an unavailable ordering record can leave the relative position unknown. Those events stay in Subagent events — relative order unavailable.

The child view shows its captured model and reasoning effort when retained. Those facts describe the child's work, even if you have since selected a different model for the parent. A model default or unavailable legacy value stays identified as such. Visible conversation and work update while the child runs. Earlier pages load retained history; interrupted live output or bounded history is labeled when it is incomplete. Assignment and result previews are separate from the actual conversation.

Counts, timestamps and history come from the host. A bounded scan can report lower-bound counts instead of exact totals. Child pages follow creation order; use Show more to reach older children, including older children working on a new assignment. Parent lifecycle activity covers the children loaded by the view and identifies bounded or incomplete history. The view keeps a bounded window of loaded children; Newest subagents returns to the newest page after older paging has moved earlier rows out of that window. Older connections can expose assignment/result inspection without the richer live conversation; the view explains that limitation and remains read-only.

Closing the tab hides the view; a transcript indicator can open it again. It stops neither the child nor its background processes. Switching parent chats withdraws the previous chat's view. Live updates do not bring an inactive tab forward. The main chat's Stop and the Processes tab's Stop controls have different targets. Processes is in the workspace's bottom Workspace utilities panel beside Terminal, separate from the Subagents view.

See independent reviews​

In Grida Desktop on macOS and Linux, choose Ultra effort in the model picker and send your request, for example “Compare the two layout drafts and tell me which assumptions matter.” Ultra work may include one independent reviewer, which the agent decides whether to start. The reviewer can read the project but does not see your chat, and its work appears in the turn's activity. Ultra is available on every supported native model connection, including subscriptions and compatible endpoints. The reviewer uses the parent's selected connection and account, with its own bounded assignment; connecting another provider does not move that work. An Ultra goal shares its inference allowance with the review. Choose how much effort to use

The reviewer gets only its assignment and permitted project-read access. It does not inherit additional folder grants, commands, writes or permission to delegate. Stop and connection shutdown also stop the review. Reopening its recorded activity starts no new request. A model's reasoning controls and provider token limits are separate from this review allowance.

Deleting a chat also deletes its subagents.

Named subagents in private clients​

A private ACP client connected to a source-built host started with --subagents can create named children and submit bounded text assignments. These client operations are separate from Desktop's read-only Subagents view. See Assign work to a named subagent for the client contract and its controls.