| name | cybernetic-parallel-subagent-workflow |
| description | Use when an approved cybernetic execution policy selects Parallel subagent-driven work with parallel-max-safe mode, explicit independence, locks, barriers, and human approval. |
Cybernetic Parallel Subagent Workflow
Preconditions
The approved plan must record parallel authority, a dependency and required-step
frontier, a wave matrix, conflict or lock ownership, failure policy, integration
barriers, and maximum concurrency.
Workflow
- Derive the next safe wave from the approved frontier; do not infer
independence from different filenames alone.
- Give each subagent one bounded domain, explicit read/write authority,
candidate identity, expected state change, and return contract.
- Prevent overlapping writes and shared mutable environment operations unless
the approved lock model serializes them.
- Collect all wave results before the integration barrier. The main agent
inspects changes, resolves conflicts, retires stale evidence, and runs the
approved integration checks.
- On a package failure, apply the approved cancel/continue policy and recompute
the next safe frontier before dispatching more work.
Test Policy
Parallel agents are not required to follow TDD or create a fixed number of
tests. Their evidence must be sufficient for the package's state-change claim;
whole-candidate Counterexample Gate remains mandatory where approved.
Boundaries
- Do not use this workflow without explicit approved parallel authority.
- Do not substitute it for serial implementer/reviewer sequencing.
- Subagents do not approve control artifacts or make final completion claims.
