| name | orx-agent-delegation |
| description | Delegate independent work to helper agent sessions with `orx agent spawn`: task selection, self-contained briefs, branch ownership, compute authorization, wakeups, and nesting or concurrency constraints. Use before spawning a helper or interpreting its result; do not delegate the literature retrieval loop. |
Delegate work to another agent
orx agent spawn creates a new top-level session in the same project. The
helper is visible to the user, receives its own worktree and transcript, and
works independently from this session.
orx agent spawn "<self-contained task>"
orx agent spawn --title "<session title>" --stdin
orx agent spawn "<task>" --harness <harness> --model <model>
orx agent spawn "<task>" --no-wake
By default, this chat resumes with the helper's closing reply. Use --no-wake
only when no follow-up is needed. A spawned session cannot spawn another helper,
and the CLI enforces the number of helpers a session may have in flight. If the
command refuses a spawn for either reason, do the work here or wait for a helper
to finish.
Choose tasks with a clean boundary
Delegate work that is genuinely independent from the node this session owns,
such as surveying an unfamiliar codebase or writing up completed results. Do
the work here when it is a step in the experiment loop already underway.
Never delegate the retrieval loop covered by orx-lit-review: the main agent
must inspect and rank the combined literature candidates itself.
Protect branches and compute
- Never give a helper a branch checked out by this session. If it must change
experiment code, tell it to create its own node and work on that node's branch.
- A frozen experiment node may not be edited by either session.
- State exactly which
orx exp run calls the helper may launch. Explicitly
forbid launches when none are authorized; otherwise the helper may infer that
the normal research loop is available.
- The helper's edits remain in its worktree. Nothing merges into this branch
automatically.
Write a standalone brief
The helper starts with an empty transcript and cannot see this conversation.
Include the project, relevant experiment and branch, metric, constraints,
allowed compute, expected output, and a concrete definition of done. Use
--stdin for a multi-paragraph brief.
