Skip to content

README.md

Latest commit

 

History

History
96 lines (67 loc) · 4.2 KB

README.md

File metadata and controls

96 lines (67 loc) · 4.2 KB
Raw

Agent Orchestration Skill Library

A Superpowers-compatible skill library for designing, building, and operating multi-agent systems. Re-architected from the original Agent Orchestration Research Library (25,000+ lines across 13 source documents) into actionable, composable skills optimized for AI agent context windows.

Architecture

Three-layer progressive disclosure:

Layer Budget Content Loaded When
Metadata ~100 tokens/skill Name + trigger description Always (skill discovery)
SKILL.md ~2,000-4,000 tokens Decision trees, gates, checklists On skill trigger
references/ Unlimited Deep knowledge, code, comparisons On demand

Each SKILL.md contains enough to make the next decision, not enough to execute it. Execution details live in references.

Skill Map

Core Decision Skills

Skill When to Use Entry Point?
formation-selection Choosing an orchestration pattern (hub-spoke, hierarchical, pipeline, mesh, swarm) Yes — start here
framework-selection Choosing an SDK/framework (Claude SDK, LangGraph, CrewAI, etc.) After formation
operational-discipline Anti-patterns, checklists, context budgets, session hygiene Cross-cutting

Infrastructure Skills

Skill When to Use
agent-communication Designing how agents share information, memory, state, consensus
agent-security Securing agent systems (8-layer defense, injection defense, sandboxing)
agent-observability Monitoring, tracing, cost tracking, evaluation

Operations Skills

Skill When to Use
agent-debugging Diagnosing and fixing agent failures (failure taxonomy, self-healing)
durable-workflows Making workflows survive failures (Temporal, checkpointing, persistence)

Frontier

Skill When to Use
cutting-edge-techniques Experimental patterns (swarms, evolution, DSPy, MoA, agent OS)

Reading Paths

Path A: First Build (deploy something this week)

  1. operational-discipline → Pre-build checklist, Single-Agent Test
  2. formation-selection → Choose pattern
  3. framework-selection → Choose SDK
  4. agent-security → Security checklist before deploy

Path B: Debugging (something is broken)

  1. agent-debugging → Classify failure, select tool
  2. agent-observability → Get trace data
  3. operational-discipline → Check anti-patterns

Path C: Production Hardening (make it reliable)

  1. operational-discipline → Context budgets, anti-patterns
  2. agent-security → Full 8-layer defense
  3. agent-observability → Monitoring and cost tracking
  4. durable-workflows → Checkpointing and recovery

Path D: Cutting Edge (what's next?)

  1. cutting-edge-techniques → Maturity gate, production-ready vs research
  2. formation-selection → Compare with standard formations
  3. framework-selection → DSPy and specialist frameworks

Superpowers Disciplines

Every skill in this library includes:

  • Hard Gates — Barriers preventing premature advancement (e.g., Single-Agent Test, security checklist)
  • Verification Checklists — Specific items to confirm before marking complete
  • Rationalization Tables — "What you might think / Why it's wrong / What to do instead"
  • Progressive Disclosure — Resources section with one-sentence descriptions of when to read each reference
  • Cross-References — Related skills with trigger conditions

Source Material

This library synthesizes:

  • Agent Orchestration Research Library (original source corpus) — 25,000+ lines, 13 documents, 8 parallel research agents
  • Superpowers — Skill-based AI agent workflow system

See research-index.md for a detailed mapping of skills to source documents.