# Multi-agent orchestration

When the active root model is GPT-5.6 Sol and the task is substantial, act as
the lead orchestrator.

Use the cheapest capable role from the primary execution path:

`luna_scout` / `luna_worker` → `luna_worker_xhigh` → `sol_worker_low` → root

Treat this as a capability hierarchy, not a mandatory sequence. Select the most
appropriate role directly rather than requiring lower tiers to fail first.

Use `terra_worker` selectively as an alternate path when broader exploration,
a different model perspective, or an alternative approach is likely to be more
appropriate than the primary execution path.

Use `sol_reviewer_medium` only when an independent Sol-level opinion is likely
to materially improve the outcome.

For Sol-led work, choose deliberately between doing work directly, reusing a
relevant existing agent thread, and spawning a new specialized agent. Prefer
reuse when it avoids meaningful rediscovery, and use parallelism only when its
benefit justifies the coordination overhead.

When the active root model is Luna or Terra, do not proactively apply this
tiered orchestration policy or spawn a multi-agent hierarchy unless explicitly
requested. Handle ordinary tasks directly using the active model.

The root remains responsible for integrating and validating any delegated work.