Follow delegated work

A subagent works on an assignment delegated by the agent in your main conversation. Use delegation for separable work, such as inspecting two parts of a project or comparing approaches before the main agent combines the findings. The main conversation remains the place to give instructions and review the combined result. A subagent is different from the agent’s response checklist, a queued follow-up, or a background shell command.

​

Request focused delegation

Start with an existing conversation in the intended project. Check its harness and model, working folder, and permissions. Describe the assignments, their boundaries, and how the results should come back together. For a read-only review of a small app, for example:

Use two subagents to review Trail Notes. Have one inspect the completion-storage design and the other inspect keyboard accessibility. Neither should edit files or install software. Combine their findings into a short implementation plan, identifying any disagreements and untested assumptions.

The active harness must expose delegation tools, and the agent must actually use them. A response saying that it plans to delegate is not evidence that a child has started. Look for subagent activity in the conversation.

​

Model choice and delegation are separate

Selecting an OpenAI, Anthropic, Google, or other provider’s model does not by itself determine the available delegation tools. Those tools come through the selected harness and its runtime configuration. Kisuke has child-lifecycle and transcript support for Codex and Claude Code. Pi can run child tasks and report nested activity, but its inspection support differs. ACP harnesses vary in the child information they expose. Do not expect every harness to provide the same child controls or a separately recorded transcript. If the agent reports that delegation is unavailable, let it perform the work in the main conversation or start another ordinary conversation with the context it needs. Rewording the same request cannot add a missing runtime tool.

​

Inspect a subagent in Zen

When a response delegates work, Zen can show named subagent chips among its activity. Running names can shimmer. The inline row shows up to three children and a count when there are more.

1. Select a subagent’s name in the conversation to open its detail in Zen’s right panel.
2. Alternatively, open Show summary, find Subagents in the environment summary, and select the child there. This section appears when the conversation has subagent activity and gives access to children beyond the inline row.
3. Read Delegated task to check what the child was actually asked to do.
4. Inspect Status and the recorded transcript. Model, Agent, and work duration appear when that information is available.
5. Return to the main conversation’s composer when you want to refine the assignment or ask for a combined conclusion.

The detail panel is a read-only inspector. Opening it does not create another conversation or send the child a message. Closing its tab only closes that view; it does not stop the child’s work.

​

Interpret the status

Where the panel groups children into Active and Done, the Done group also contains failed and stopped children. It is a list of ended executions, not a list of verified successes. A completed launch tool can mean only that a background child was started. The child’s status and the main response can finish at different times. Inspect remaining children even after the parent has posted an answer.

​

Read the child’s own work

A Codex child can include inherited context. When the inspector shows Inherited from parent, expand it to read that context separately from the child’s own execution. The Worked for value, when present, is calculated from recorded child turns and excludes inherited parent turns. It is not a measure of the quality or completeness of the result, and the panel does not show a continuously updated last-activity age. The transcript is loaded when you open the child and refreshed when its status changes. It is not a continuous live stream of every child token. While a child remains Working, the visible transcript can lag behind its current work.

​

Coordinate changes to project files

Do not use the number of subagents as evidence that each has an isolated checkout. The inspector itself creates no worktree or file copy. In particular, Pi child tasks use the parent’s working directory. For parallel implementation, make file ownership explicit in the request and ask the main agent to reconcile the results. Use worktrees when you need separate working folders, and verify the actual folder before work begins. After delegated edits, inspect Git Review and the reported checks. A Done status does not establish that tests passed, changes were committed, or another child’s edits were preserved.

​

Use the phone summary

The phone’s chat summary can list subagent names and the same status vocabulary. That list is informational; it does not provide the desktop Zen child-transcript inspector described above. For detailed child history, open the parent conversation in desktop Zen when the harness has recorded a supported transcript. Keep the computer that owns the conversation connected. See connected devices if other project tools are also unavailable.

​

Recover from missing or stale details

For a separate discussion that you control directly, use Side Chat. For an objective that continues across responses, use a goal. Neither is the same as inspecting delegated child work.