agentic-parallelism-decision
This skill should be used when the user asks 'how do I parallelize this', 'which parallel primitive should I use', 'subagent vs agent teams vs workflow', 'should I use a worktree', 'run work in the background', or needs to choose between subagents, agent view, agent teams, dynamic workflows, and git worktrees. Routes to the right primitive skill; it does not re-explain each one.
Agentic Parallelism Decision
Entry point for "how do I parallelize this?". Classify the intent, then route to the owning primitive skill. This skill routes; it does not duplicate the primitives.
A primitive = a Claude-native parallelism mechanism (subagent, agent view, agent teams, dynamic workflow, git worktree).
Decision (fast path)
Table evidence: [DOC] (each row verified against the owning primitive's skill) except the fallback row, which is [INFERENCIA].
| If the need is… | Use | Owner skill |
|---|---|---|
| A few delegated side tasks inside one session, isolated context | Subagents (Task tool) | agent-creator, katas-hub-and-spoke-isolation |
| Independent background sessions you dispatch and monitor | Agent view | agent-view-ops |
| Teammates that message each other on a shared task list | Agent teams (experimental) | agent-team-orchestration |
| Script-held fan-out/cross-check across many agents, resumable | Dynamic workflow | dynamic-workflow-forge |
| File isolation so parallel edits don't collide | Git worktree | worktree-isolation |
Classifier
- Do parallel branches edit the same files? Yes → add
worktree-isolation(compose it with any of the below). No → keep going. [DOC] - Do branches need to talk to each other? Require real peer messaging, not just
parallel read-only lookups. Yes →
agent-team-orchestration(experimental). No → degrade to subagents; don't route to teams for independent branches. [DOC] - Should branches run as detached background sessions you monitor? Yes →
agent-view-ops. No → keep going. [DOC] - Is the orchestration large, repeatable, and resumable? Yes →
dynamic-workflow-forge. No → keep going. [DOC] - Otherwise: a few isolated side tasks → subagents (
agent-creator). [INFERENCIA]
Full trade-off table (cost, coordination, isolation, resumability, peer messaging,
file-conflict safety): references/parallelism-selection-matrix.md.
Guardrails
- Prefer the least powerful primitive that fits — subagents before workflows before teams. [INFERENCIA]
- Worktrees are isolation, not coordination; compose, don't substitute. [DOC]
- Treat agent teams as experimental/feature-gated; confirm availability first.
- Agent view is research-preview and version-gated (v2.1.139+); confirm version before asserting availability. [DOC]
- Propagate structured errors across branches (
katas-multiagent-error-propagation).
Related
- Primitives:
agent-creator,agent-view-ops,agent-team-orchestration,dynamic-workflow-forge,worktree-isolation. - Patterns:
katas-hub-and-spoke-isolation,katas-multiagent-error-propagation,katas-deterministic-agent-loop,katas-multipass-prompt-chaining.
Contract
- Aceptación: clasifica la intención y rutea a UN primitivo (subagente/agent-view/team/workflow/worktree) con rationale. [EXPLICIT]
- Límites: solo rutea; NO re-explica ni implementa el primitivo. [EXPLICIT]
- Casos borde: señales mixtas → presentar trade-offs y dejar decidir, no adivinar. [EXPLICIT]
- Supuestos: los 5 primitivos disponibles según probe. [SUPUESTO]
- Trade-off: ruteo centralizado evita duplicación pero añade un salto de indirección. [EXPLICIT]
Packet
Capas del packet, cargables bajo demanda (disciplina ICM: una capa por vez, nunca todas juntas): references/ guías de profundidad (cargar UNA por etapa) · assets/ recursos estáticos.
- Repository
- JaviMontano/claude-plugins
- Commit
e8f986b
- Last updated
- First committed
Is this your skill?
If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.