subagent-dispatch

Guidance for splitting independent coding tasks among multiple AI agents working in parallel. It covers when tasks can safely run at the same time and how to write focused instructions for each agent.

In plain words
What is it for?
Use it to design parallel agent workflows and create isolated prompts for separate problem areas.
Why use it?
It reduces duplicated work and conflicts by checking dependencies, shared files, and available agent capacity before dispatching tasks.

Skill for Claude CodeCodex

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add skills/auerbachb/claude-code-config/subagent-dispatch
Any agent
npx skills add auerbachb/claude-code-config --skill subagent-dispatch
Clone the repo
git clone --depth 1 https://github.com/auerbachb/claude-code-config

Made for: Claude Code, Codex.

Per session 46 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 2,445 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00046 $0.02445
Opus 5 $0.00023 $0.01222
Sonnet 5 $0.00009 $0.00489
Haiku 4.5 $0.00005 $0.00245

Measured 2d ago against content hash 50116584df7b, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

subagent-dispatch scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

.claude/skills/subagent-dispatch/SKILL.md · 194 lines

How it starts

The opening of the file, as written. The whole thing — 194 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Core principle: one agent per independent problem domain.

This skill teaches the craft of designing and writing independent parallel-agent prompts. Spawn mechanics, model selection, and Phase A/B/C orchestration are in .claude/rules/subagent-orchestration.md — do not duplicate them here.


Step 1: Decide Whether to Parallelize

Run this decision tree before dispatching anything.

Multiple tasks?
 └─ Are failures/tasks related?
     ├─ YES (related) → Single agent investigates all
     └─ NO (independent) → Can they work without shared state?
         ├─ NO (shared state, e.g. same file) → Sequential agents
         └─ YES (no shared state) → Parallel dispatch

Parallelize when:

  • Tasks span different subsystems, files, or problem domains
  • Each task can be understood without context from the others
  • Agents will not edit the same files or shared state
  • The ceiling allows it (see Step 2)

Do NOT parallelize when:

  • Failures are related (fixing one may fix others — investigate first)
  • Agents would write to the same files, branch, or session state
  • You don't yet know what's broken (explore first)
  • The pipeline ceiling is already full

This repo's vocabulary: /wave calls this the "dependency- and overlap-filtered independent set"; /subagent calls it "disjoint chains run as parallel pipelines." Both terms mean the same thing as "independent problem domain."


Step 2: Check the Pipeline Ceiling Before Dispatching

Pipeline ceiling (3–4 active PRs, per .claude/rules/subagent-orchestration.md) — how many concurrently reviewed PRs the parent thread manages.

Silence/bgwork ceiling (bgwork-ceiling.sh) — a separate backstop that fires when background work goes quiet. Do not confuse the two.

Count your currently open PRs (gh pr list --author "@me" --state open) as a conservative proxy for pipeline load. Dispatch only enough agents to stay at or below the pipeline ceiling (3–4). Agents that would exceed the ceiling queue inline — they do not get their own chips. (The canonical ceiling definition — actively CR-polled PRs authored by @me — lives in .claude/rules/subagent-orchestration.md; dormant open PRs are counted here for safety.)

Read the full file on GitHub · 194 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 2d ago First seen · 194 lines · 46 tokens per session scan A 50116584df7b

Subscribe to this mod's changes

subagent-dispatch is a skill published in the GitHub repository auerbachb/claude-code-config (5 stars, last pushed 3d ago), licensed MIT. It adds 46 tokens to every session and 2,445 once invoked, about $0.0002 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.

Related

Other skills, from other repositories

implement

Canonical implementation parent command. Executes the active track task-by-task using TDD and verification gates, and routes to status, coverage, or revert when the user asks for progress, measurement, or rollback explicitly.

drafthq/draft · 43 tokens

learn

Scan codebase to discover coding patterns and update draft/guardrails.md. Learns conventions (skip in future) and anti-patterns (always flag). Supports migration from workflow.md and pattern promotion.

drafthq/draft · 42 tokens

adapt-to-project

Use this skill to diagnose and improve an adopter repository's agent guidance or to walk through the four classes of post-install change (substitution, .upstream companion merges, discovery + restructuring, within-layout consolidation). Repository anchoring is marker-independent and read-only by default. Post-install…

eugenelim/agent-ready-repo · 103 tokens

new-skill

Scaffold a new logic- skill in the Logic-Lens repo and wire it into every place a skill must be registered, so no step is missed. Use when adding a seventh (or later) skill to Logic-Lens.

hyhmrright/logic-lens · 50 tokens

compile-okf

Use when a catalogue maintainer needs to compile declared OKF authoring bundles into generated router and reviewed procedure Skills, or check that committed generated output is current.

eugenelim/agent-ready-repo · 36 tokens

sol-governed-workers

Use when Codex work should route bounded implementation to lower-cost workers and reserve Sol for risky plan review or final evidence gates, or when installing or validating this workflow.

BusyBee3333/sol-governed-codex · 39 tokens