Agent Orchestration

SkillProductivity

Multi-agent task orchestration for Claude Code. When given multiple tasks to execute, use this skill to spawn parallel worker agents through Sibyl's MetaOrchestrator instead of Claude Code's built-in Task tool. This enables true multi-process execution with budget controls, quality gates, and coordination.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Agent Orchestration skill

What this skill tells your AI

The instructions your AI receives, as published by diegosouzapw/awesome-omni-skill in skills/data-ai/agent-orchestrate/SKILL.md and read by ahel’s review.

CRITICAL: This skill teaches you how to dispatch tasks to multiple worker agents.

When you're asked to work on multiple tasks, or to "dispatch agents", "parallelize work", or "orchestrate" - use Sibyl's MetaOrchestrator via the CLI, NOT Claude Code's built-in Task tool.

Why This Matters

Claude Code's built-in Task tool spawns inline subagents within YOUR process. This is useful for research, but NOT for parallel task execution because:

  • Subagents run sequentially, not in parallel
  • No budget controls or quality gates
  • No coordination between workers
  • Context bloat as results accumulate

Sibyl's MetaOrchestrator spawns separate Claude processes that:

  • Run truly in parallel (up to N concurrent)
  • Have isolated worktrees (no git conflicts)
  • Go through implement → review → rework loops
  • Track budget and can pause when limits hit
  • Report progress through the message bus

Quick Start

# One-liner: dispatch all todo tasks with 3 parallel agents
sibyl agent dispatch --status todo --strategy parallel --max 3

# With budget limit
sibyl agent dispatch --status todo --budget 25.00 --gates lint,test

# Specific tasks
sibyl agent dispatch task_abc123 task_def456 --strategy parallel

The Dispatch Command

sibyl agent dispatch is the high-level command that handles everything:

sibyl agent dispatch [TASK_IDS...] [OPTIONS]

Options:
  --status, -s    Filter tasks by status (todo, doing, blocked)
  --priority      Filter tasks by priority
  --strategy      Execution strategy: parallel, sequential, priority
  --max, -m       Max concurrent agents (default: 3)
  --budget, -b    Budget limit in USD
  --gates, -g     Quality gates (comma-separated: lint,test,typecheck)
  --dry-run       Show what would be dispatched
  --project, -p   Project ID (auto-resolves from cwd)

Examples

# Dispatch all todo tasks
sibyl agent dispatch --status todo

# Dispatch high-priority tasks only
sibyl agent dispatch --status todo --priority high,critical

# Parallel with 4 workers and $50 budget
sibyl agent dispatch --status todo --strategy parallel --max 4 --budget 50.00

# Sequential with quality gates
sibyl agent dispatch task_abc --strategy sequential --gates lint,test,typecheck

# Dry run first
sibyl agent dispatch --status todo --dry-run

Lower-Level Commands

For fine-grained control, use the sibyl agent orchestrate subcommands:

# Initialize orchestrator
sibyl agent orchestrate init

# Queue specific tasks
sibyl agent orchestrate queue task_abc123 task_def456

# Set strategy
sibyl agent orchestrate strategy parallel --max 4

# Set budget
sibyl agent orchestrate budget 50.00 --alert 0.8

# Start processing
sibyl agent orchestrate start --gates lint,test

# Monitor progress
sibyl agent orchestrate status

# Pause/resume
sibyl agent orchestrate pause
sibyl agent orchestrate resume

Strategies

StrategyBehaviorUse When
parallelRun up to N tasks concurrentlyIndependent tasks, need speed
sequentialOne task at a timeDependent tasks, shared state
priorityOne at a time, highest priority firstMixed urgency, limited budget

Quality Gates

Quality gates run after each implementation cycle:

  • lint - Run linter (ruff, eslint, etc.)
  • test - Run test suite
  • typecheck - Run type checker (pyright, tsc)

If any gate fails, the worker enters a rework cycle (up to 3 attempts).

sibyl agent dispatch --status todo --gates lint,test,typecheck

Monitoring

# Check orchestration status
sibyl agent orchestrate status

# Output shows:
#   - Queue size and active workers
#   - Completed/failed counts
#   - Budget utilization
#   - Rework cycles

Workflow: Orchestrating a Sprint

# 1. See what's available
sibyl task list --status todo

# 2. Preview what would be dispatched
sibyl agent dispatch --status todo --dry-run

# 3. Launch with controls
sibyl agent dispatch --status todo \
  --strategy parallel \
  --max 3 \
  --budget 30.00 \
  --gates lint,test

# 4. Monitor progress
sibyl agent orchestrate status

# 5. Pause if needed
sibyl agent orchestrate pause

When to Use What

Use Sibyl orchestration (sibyl agent dispatch) when:

  • You need multiple implementation tasks done in parallel
  • Tasks are independent and can run concurrently
  • You want budget controls and quality gates
  • You're orchestrating a sprint or batch of work

Use Claude Code's built-in Task tool when:

  • Researching or exploring code (reading files, searching)
  • Asking specialized agents questions
  • Delegating subtasks within a single implementation

Key insight: These are complementary, not exclusive!

Worker agents spawned by Sibyl orchestration are encouraged to use their built-in Task tool for research, exploration, and subtask delegation. The hierarchy is:

MetaOrchestrator (Sibyl)
├── Spawns parallel worker agents for independent TASKS
│
Worker Agent 1                    Worker Agent 2
├── Uses Task tool for research   ├── Uses Task tool for research
├── Spawns subagents as needed    ├── Spawns subagents as needed
└── Implements assigned task      └── Implements assigned task

The rule is simple:

  • Cross-task parallelism → Sibyl orchestration (separate processes, isolated worktrees)
  • Within-task work → Built-in Task tool (subagents, research, exploration)

Architecture

MetaOrchestrator (Tier 1)
├── Manages sprint/project-level coordination
├── Maintains task queue
├── Spawns TaskOrchestrators per strategy
└── Tracks budget and metrics

TaskOrchestrator (Tier 2)
├── Manages single task lifecycle
├── Runs implement → review → rework loop
├── Enforces quality gates
└── Reports to MetaOrchestrator

Worker Agent (Tier 3)
├── Actual Claude process
├── Has isolated git worktree
├── Implements the code
└── Communicates via message bus

Troubleshooting

"No tasks to dispatch"

  • Ensure you have tasks with matching status/priority
  • Check project context: sibyl context
  • Try --dry-run to see what matches

"No orchestrator or project specified"

  • Link your directory to a project: sibyl project link <project_id>
  • Or specify explicitly: --project proj_abc123

Budget exhausted

  • Check status: sibyl agent orchestrate status
  • Increase budget: sibyl agent orchestrate budget 100.00
  • Resume: sibyl agent orchestrate resume

Signals

GitHub stars
57
Forks
19
Last commit
Mar 2026
Advanced
Catalog kind
skill
Gateway key
agent-orchestrate
Source
github.com/diegosouzapw/awesome-omni-skill