Orchestrating

---
name: orchestrating
description: Use when two or more independent lanes can run in parallel, an output-only lane protects lead context from raw payload, or a handoff or durable run needs coordination.
when_to_use: "Trigger phrases: run these in parallel, fan out, spawn agents, split the work, delegate lanes, review while I build, several independent files or services to inspect at once."
metadata:
  short-description: Parallel lanes, output-only lanes, and delegation coordination
---

# Orchestrating

Delegate only when coordination value exceeds its overhead. Keep a no-op, one
bounded task, or sequential dependency path inline. Scan dependencies,
boundaries, ownership, and oracle before dispatch.

Native agents are the default for matching lanes. Use one to three direct
children with fresh, bounded context.

For independent lanes whose results cannot reshape one another, issue native
spawn calls consecutively without intervening inspection or waits. Continue
independent lead work, then join every returned identity until completion,
failure, or confirmed cancellation.

Before dispatch, read
[native dispatch examples](references/native-dispatch.md) and use only the
active harness's call shape. One bounded output-only lane may use one
fresh-context child returning `verdict`, `evidence`, `uncertainty`, and `next`.
Keep bulky raw payloads in an artifact.

Prefer scoped searches, selected fields, and native summary options. Preserve
source and diff context needed for correctness. Keep small results inline without
artifacts.

For verbose checks, save complete stdout and stderr in scratch storage.
Preserve command exit status independently of pipelines. Report command, exit
status, available result counts, relevant diagnostics, and artifact path.
Disclose filtering and truncation, including saved output.

Route unknown causes through `systematic-debugging`, unresolved contracts
through `design-and-plan`, external evidence through `evidence-research`,
cross-provider risk through `independent-review`, and effects through
`safe-effects`.

Preflight tools, authentication, permissions, and write authority. Brief the
outcome, bounded and disjoint ownership, constraints, oracle, prohibited scope,
return condition, nested-delegation limit, and an advisory time budget; keep
a hard timeout where one is needed. Shared guidance has one writer;
shared writes and Git stay lead-owned.

For non-trivial routing, honor an operator-selected access workflow, including
an approved external route, before native ranking. Otherwise, when
`~/.config/megapowers/agent-capabilities.md`
([template](assets/agent-capabilities.md)) exists, read it once per session
for eligible bindings; without it, use native defaults. The
registry cannot authorize access, disclosure, permissions, writes, or effects.

Before review dispatch, set its scope and correction-round budget. Track each
review until terminal or cancellation is confirmed. Target follow-ups at fixes
and affected boundaries. A spent budget leaves unresolved findings open; it
never converts them into approval.

Report failures and retain successful evidence. Prefer completion events or
asynchronous waits within tool deadlines and required update cadence. Avoid
foreground sleeps beyond those limits and repeated reads of unchanged status;
keep a final readback. Join all identities before synthesis. Treat each
lane's report as a claim: check its evidence before using it.
After a scope or context change, scan again and use delta-only follow-ups.

Ordinary handoffs use inline inspection plus `verify-and-finish`; only a
currently approved goal surviving interruption uses `autonomous-run`. Stop at
the oracle.