Skip to main content

Delegation

ExperimentalAvailable on: WindowsPresent in the build, unproven. May change or be removed.

Assigning work to another agent, and reading the difference between assigned, queued, and started.

Delegation creates an edge assigning a task to another agent. It is the mechanism behind multi-agent work, and the place where optimistic status reporting does the most damage.

Delegating

From the TUI:

/delegate

Related commands: /agents lists the fleet, /ask poses a question without assigning work, /review requests a review, /send messages an agent directly.

NOTE

/ask and /delegate differ in commitment. Asking gets you an answer; delegating creates durable work that someone has to finish or cancel.

Assigned is not started

A delegation edge exists the moment you create it. That says nothing about whether the receiving agent began.

QuestionAnswered by
Was the task assigned?The delegation edge
Did the message reach the agent?Delivery state
Did a turn start?received_and_started specifically
Is the work done?Task state

All four can disagree. See delivery lifecycle for the six real delivery states, including needs_enter — text staged in a composer that was never submitted, which looks exactly like a working agent from the outside.

Isolation

Delegated work should run in its own git worktree so parallel agents do not collide in the same files. See worktrees.

IMPORTANT

A worktree is file isolation, not a security boundary. A delegated agent can still reach the rest of your machine and any credentials in its environment.

Supervising delegated work

The failure mode to guard against is a fleet that looks busy and is not.

nala tasks list

Read task state rather than inferring from terminal activity or the number of panes with output scrolling.

WARNING

FleetView can show more agents live than actually exist. Confirm against task state before concluding work is in progress. See known limitations.

Verification is not delegated

The delegating agent — or you — still has to verify. Agent-reported success is not evidence. Read the diff, run the tests.

Where a task carries a verification gate, the evidence has to exist before it completes. Where it does not, you are the gate.

When not to delegate

  • Work you cannot verify. Delegating something you cannot check produces confident output and no way to trust it.
  • Work that touches the same files as another agent's task. Split it differently or run it in sequence.
  • Exploratory work. Delegation suits scoped tasks; open-ended investigation is usually faster in one session.