Parallel Agent Work Planner
Plan concurrent coding-agent work with explicit dependencies, exclusive workspace ownership, a bounded worker count, evidence requirements, and an ordered integration path.
Capabilities
- Finds hidden dependencies and serialization points before dispatch
- Defines independent work packets with exclusive writable ownership
- Sets concurrency from review and infrastructure capacity
- Requires inspectable evidence from every worker
- Plans dependency-aware integration and final-head verification
SKILL.md
---
name: parallel-agent-work-planner
description: Plan parallel coding-agent work when a task may benefit from concurrent agents, worktrees, or subagents. Produce a dependency-aware work packet plan; do not dispatch agents or change files.
metadata:
author: Ryan VerWey
reviewed: "2026-09-10"
---
# Parallel Agent Work Planner
Turn a software task into a safe parallel execution plan. Use concurrency only for work that can proceed independently and be reviewed on its own.
## Inspect
1. Read the goal, acceptance criteria, repository instructions, and current revision.
2. Identify shared contracts, generated artifacts, migrations, lockfiles, deployment state, and other serialization points.
3. Draw the task dependency graph before assigning workers.
4. Estimate review capacity and shared infrastructure limits.
## Decide
Keep the task sequential when it is small, tightly coupled, destructive, or dominated by one shared boundary. Parallelize only packets that have stable inputs, distinct writable ownership, independent checks, and no need for another packet's unreviewed output.
Use read-only research or review packets when exploration can proceed independently. Do not invent work merely to occupy another agent.
## Define Each Work Packet
For every packet, provide:
- goal and observable output
- prerequisites and dependency edges
- exact repository and base revision
- writable files or boundaries
- relevant read-only context
- non-goals and forbidden external actions
- required checks and evidence
- stop conditions and blocker reporting
- integration order and result handoff
Assign one writable owner per file or shared contract. Workers may inspect neighboring code but must not refactor, publish, merge, deploy, or close tracking work unless the user separately grants that authority.
## Set the Concurrency Budget
Recommend the smallest worker count that shortens the critical path. Limit it by independent packet count, reviewer capacity, CI and sandbox capacity, rate limits, and expected coordination cost.
Explain which packets run concurrently and which remain ordered. If independence is uncertain, keep the edge sequential and state what evidence would make fan-out safe.
## Plan Fan-In
Name one integration owner. Specify the order for contracts, consumers, tests, documentation, generated files, review fixes, and release metadata.
Require every worker to return its base revision, changed files or research scope, commands and outcomes, source URLs, assumptions, blockers, and interaction risks. Require focused checks after each integration step and the full project checks on the combined head.
Add an independent review of the combined diff. Passing checks on separate branches are not evidence that the integrated result works.
## Return
Produce:
1. Parallelism decision and rationale.
2. Dependency graph in a compact text or table form.
3. Work packets with exclusive ownership and checks.
4. Concurrency budget and dispatch groups.
5. Fan-in order and conflict policy.
6. Final verification and release gates.
7. Assumptions, blockers, and work kept sequential.
## Sources
Reviewed September 10, 2026. Product capabilities below are examples; the planning method is SLOPSTACK guidance.
- [GitHub Copilot app agent sessions](https://docs.github.com/en/copilot/how-tos/github-copilot-app/agent-sessions)
- [GitHub Copilot SDK custom agents and sub-agent orchestration](https://docs.github.com/en/copilot/how-tos/copilot-sdk/features/custom-agents)
- [GitHub Copilot app slash commands](https://docs.github.com/en/copilot/reference/github-copilot-app-reference/slash-commands)
- [OpenAI model guidance](https://developers.openai.com/api/docs/guides/latest-model)